[codex] fix(subtitles): stream live transcribe_translate cues (#177)

* fix(subtitles): stream live transcribe_translate cues * fix(subtitles): harden live AI transcription
2026-06-18 12:16:27 -04:00
parent 6e1f79e8a9
commit b3198276f7
10 changed files with 1590 additions and 109 deletions
@@ -0,0 +1,311 @@
+# Plan: interleave `transcribe_translate` so live AI subtitles follow the playhead
+
+Issue: https://github.com/Silo-Server/silo-server/issues/154
+Area: `internal/subtitles/ai` (+ `internal/playback` extraction, `internal/config`)
+Type: bug fix to an existing (shipped) capability — makes the chained AI-subtitle
+path deliver its designed live behavior. No new client-facing contract.
+
+Commands assume the repository root is the cwd.
+
+## Problem (confirmed)
+
+`Service.runTranscribe` (`internal/subtitles/ai/service.go`) runs the chained
+`transcribe_translate` kind as two **global, sequential** stages:
+
+1. Transcribe the **entire** file (progress 5%→70%). Transcript cues are not
+   streamed for the chained kind —
+   `streamTranscript := streaming && job.Kind == JobKindTranscribe`
+   (`service.go:417`), so it is `false` for `transcribe_translate`.
+2. Only then translate **all** cues (70%→95%), streaming each translated batch
+   as it lands (`service.go:482-493`).
+
+Because translation cannot start until step 1 finishes for the whole file, the
+first translated cue streams minutes into the job, long past the playhead.
+
+### Two distinct bottlenecks (both must be addressed)
+
+- **Transcription stage is global.** Translation waits for the whole-file ASR
+  pass. This is the dominant term for a feature film (~20 min ASR) and the
+  primary cause. Fixed by interleaving translation per chunk (Phase 1).
+- **Extraction is global and precedes any callback.** `Transcribe` calls
+  `t.extract(...)` (`transcriber.go:124`) before the `onChunk` loop, and
+  `ExtractAudioChunks` blocks on a single whole-track ffmpeg pass
+  (`internal/playback/audio_extract.go:34,65`). So even with per-chunk
+  interleave and small chunks, the first cue still waits for full-file
+  extraction (~1–2 min for a 2h film). Fixed by playhead-first incremental
+  extraction (Phase 2).
+
+The transcriber already exposes the per-chunk seam: `Transcribe` invokes
+`onChunk(cues, done, total)` per chunk, playhead-first (`transcriber.go:111-174`),
+and plain `transcribe` already streams through it (`service.go:426-429`).
+
+Secondary defect found while confirming: `WhisperTranscriber.SetExtraction`
+clamps any `chunkSeconds` outside `[60, 600]` to the **default 600**
+(`transcriber.go:100-102`), so a small live chunk value silently becomes 10 min
+instead of clamping to the floor.
+
+## Goal & honest latency targets
+
+Translated cues stream playhead-first for a session-attached
+`transcribe_translate` job, instead of waiting for whole-file transcription.
+
+Time-to-first-translated-cue, by phase:
+
+- **Today:** `extract(full) + transcribe(full) + translate(first batch)` — many
+  minutes (≈ full ASR pass).
+- **After Phase 1 (interleave, existing full extraction):**
+  `extract(full) + transcribe(first chunk) + translate(first chunk)` — removes
+  the whole-file ASR pass from the critical path; ≈ extraction time (~1–2 min for
+  a 2h film), then cues follow the playhead. Already usable for live after a
+  one-time wait.
+- **After Phase 2 (playhead-first incremental extraction):**
+  `extract(one chunk near playhead) + transcribe + translate` — single-digit
+  seconds. This is what the issue means by "within seconds."
+
+Phase 1 delivers most of the value with low risk; Phase 2 closes the remaining
+extraction latency. Ship Phase 1 first; do not claim "within seconds" until
+Phase 2 lands.
+
+### Non-goals (this plan)
+
+- Pacing/throttling to the playhead (rolling window). Optional in the issue;
+  follow-up below.
+- Cross-chunk translation-context continuity beyond batch context neighbours.
+
+## Decision: interleave only when streaming
+
+**Session-attached (live) jobs** use the interleaved per-chunk path.
+**Background (non-streaming) `transcribe_translate` jobs keep the current
+whole-file transcribe-then-translate-at-end path unchanged.** Rationale: the
+issue is live-only; this preserves background output byte-for-byte (full
+cross-file translation context, no incremental-extraction cost) and keeps the
+two behaviors cleanly separated by the existing `streaming` boolean. The
+shared transcript-storage step is factored so it is not duplicated.
+
+---
+
+## Phase 1 — interleave translation per chunk (streaming jobs)
+
+### 1a. `internal/subtitles/ai/service.go` — `runTranscribe`
+
+For the chained kind **when `streaming`**, replace the kind-specific `onChunk`
+body (`service.go:423-430`) with per-chunk translate-and-stream:
+
+- Capture a translate error in a closure: `var translateErr error` and an
+  accumulator `var translatedAccum []SubtitleCue`.
+- In `onChunk(chunk, done, total)`:
+  - If `translateErr != nil` or `len(chunk) == 0`, return (skip silence; stop
+    translating after a prior failure). Do **not** abort transcription.
+  - Else call `s.translator.Translate` on **that chunk's** cues. On error: set
+    `translateErr`, emit `TranslationFailed` once, and return (let transcription
+    continue). On success: stream translated cues via `TranslationCues` and
+    append to `translatedAccum`.
+  - Progress: single 5%→95% band keyed on chunk `done/total`
+    (message "Transcribing & translating").
+- Keep `onChunk` signature unchanged (no error return). Aborting on translate
+  failure via the callback is intentionally avoided — see 1b.
+
+After `Transcribe` returns:
+
+- **Always store the transcript track** (unchanged, `service.go:447-466`,
+  including the transcript `SubtitleReady` broadcast). This must run regardless
+  of `translateErr`.
+- If `translateErr != nil`: `finishWithError(ctx, job, translateErr)`. The
+  transcript is already persisted (cache preserved), and the live session
+  already received `TranslationFailed`.
+- Else: assemble the translated track from `translatedAccum` (`sortCuesByStart`
+  + `StoreSubtitle`, `service.go:499-514`), `CompleteJob`, and
+  `TranslationCompleted` (`service.go:515-522`).
+- Remove the whole-file translate-at-end (`service.go:479-497`) for the
+  streaming path only.
+
+Non-streaming chained jobs fall through to the existing whole-file path,
+untouched.
+
+### 1b. Transcript-on-translation-failure (resolves review P1)
+
+Today the full transcript is stored before translation starts
+(`service.go:447` then `:482`), so a translation failure still leaves a
+transcript/cache row and an already-broadcast `SubtitleReady`. The interleaved
+path preserves this exactly because:
+
+- a per-chunk translate failure sets `translateErr` but lets the ASR pass run to
+  completion (it does **not** return an error from `onChunk`), so `Transcribe`
+  returns the full cue set normally;
+- the transcript is then stored on the same post-transcribe path as today,
+  before the job is failed for the translated track.
+
+An ASR failure or context cancellation still returns an error from `Transcribe`
+and yields no transcript — identical to today.
+
+### 1c. Live progress semantics (resolves review P2)
+
+`SubtitleTranslationCuesPayload.Done/Total` is documented as overall progress
+(`internal/playback/realtime.go:172-181`). For live `transcribe_translate`:
+
+- `TranslationCues` carries **`done = chunks completed, total = total chunk
+  count`** — the same chunk-granular convention plain `transcribe` already emits
+  (`service.go:427-428`). Monotonic and stable across chunks.
+- `TranslationStarted.TotalCues = 0` (indeterminate), consistent with the
+  transcribe path today (`service.go:412`), since the cue total is unknown before
+  ASR completes.
+- Cues carry absolute Start/End, so the client places them regardless of
+  arrival order; unpause is driven by cue arrival near the playhead, not by
+  `Done/Total`. Confirm with client teams that `Done/Total` is treated as a
+  fraction, not absolute cue indices (clients already receive chunk counts from
+  the plain `transcribe` path, so this should hold).
+
+### 1d. Small live chunks + clamp fix
+
+- Add `TranscribeJobRequest.ChunkSeconds int` (0 = configured default). In
+  `Transcribe`, use `req.ChunkSeconds` when `> 0`, else the atomic default
+  (`transcriber.go:121-122`); validate/clamp identically to `SetExtraction`.
+- Fix the clamp in `SetExtraction` (`transcriber.go:99-105`): clamp
+  out-of-range to the nearest bound, not to the 600 default.
+- Lower `minASRChunkSeconds` from 60 to 15 (`transcriber.go:26`); update the
+  doc comment about request count / boundary word-clip tradeoff.
+- Config: add `subtitle_ai.live_asr_chunk_seconds` (default 30):
+  - `internal/config/config.go`: new field beside `ASRChunkSeconds`
+    (`config.go:274-277`).
+  - `internal/config/db_loader.go`: load it next to `asr_chunk_seconds`
+    (`db_loader.go:472-476`).
+  - `internal/subtitles/ai/config.go`: add `LiveASRChunkSeconds int` to the
+    service `Config` so the service can set `req.ChunkSeconds` for streaming jobs.
+  - `internal/api/router.go`: plumb through `effectiveSubtitleAIConfig` and the
+    `OnConfigChange` `UpdateConfig` (`router.go:1024-1055`).
+
+### 1e. Phase 1 tests
+
+- `service_test.go`: fake `Transcriber` emitting several chunks playhead-first
+  via `onChunk`; fake `Translator` recording calls. For a streaming
+  `transcribe_translate` job assert: one `Translate` per non-empty chunk; first
+  `TranslationCues` streamed before the last chunk is transcribed; empty chunks
+  skipped; final stored translated track = all translated cues sorted by start;
+  `Done/Total` is chunk-granular and monotonic.
+- Failure case: a translate error on chunk N sets the failure but transcription
+  completes, the transcript track **is** stored, `TranslationFailed` is emitted,
+  and the job ends via `finishWithError`.
+- Non-streaming `transcribe_translate`: unchanged whole-file path; correct final
+  track; no `TranslationCues`.
+- `transcriber_test.go`: `SetExtraction` clamp fix and `ChunkSeconds` override
+  (incl. the new 15s floor).
+
+---
+
+## Phase 2 — playhead-first incremental extraction (true seconds latency)
+
+Phase 1 still waits on full-file extraction before the first cue. Phase 2
+removes that by extracting and consuming chunks incrementally, seeked to the
+playhead.
+
+### 2a. New incremental extractor in `internal/playback`
+
+Add a streaming, seek-based extractor alongside `ExtractAudioChunks`:
+
+```
+ExtractAudioChunksFrom(ctx, filePath, audioTrackIndex, dir, ffmpegPath,
+    startSec float64, chunkSeconds int,
+    onSegment func(AudioChunk) error) error
+```
+
+- Runs one ffmpeg pass with `-ss <startSec> -i <file> ... -f segment
+  -segment_list segments.csv` covering `startSec`→end.
+- Consumes segments **as they complete**: poll the segment-list CSV (ffmpeg
+  appends `filename,start,end` when it closes each segment) on a short interval;
+  for each new row, invoke `onSegment(AudioChunk{Path, Start: startSec + csvStart})`.
+  Reconcile any trailing rows after `cmd.Wait()` so the final segment is not
+  missed.
+- Honors `ctx` (CommandContext kills ffmpeg on cancel) and `onSegment` errors
+  (stop the pass). Caller owns `dir`; segments are deleted by the consumer after
+  ASR to cap disk at one extraction.
+- Preserves exact per-segment starts from the CSV, so timing does not accumulate
+  drift within a pass (only the initial `-ss` seek is packet-accurate, sub-second,
+  and is further corrected by the existing `ProbeAudioStartOffset`).
+
+Rejected alternative — **per-chunk seek extraction** (`-ss s -t d` per chunk):
+simpler control flow but O(n) ffmpeg spawns and packet-aligned per-chunk seeks
+that can clip words at every seam. The single-pass streaming consumer keeps the
+existing segment-muxer boundary quality and far fewer process spawns.
+
+### 2b. Transcriber incremental mode
+
+- Add `TranscribeJobRequest.Incremental bool`, set by the service for streaming
+  jobs.
+- When `Incremental`, `Transcribe` drives playhead-first in two streaming passes
+  (reproducing today's `chunkOrderForPosition` order): pass A `startSec =
+  playhead → end`, then pass B `0 → playhead`. Each completed segment is fed
+  straight into the existing per-chunk ASR loop body (transcribe → `onChunk`),
+  which Phase 1 already wires to translate+stream. The full cue set is still
+  accumulated and returned for transcript storage.
+- When not `Incremental` (background, and plain transcribe unless we opt it in),
+  keep the existing whole-file `ExtractAudioChunks` + `chunkOrderForPosition`
+  path.
+
+Note: timing composition (`-ss` reset PTS + segment CSV start + probe offset)
+must be verified with a fixture; see tests.
+
+### 2c. Phase 2 tests
+
+- `audio_extract` test (or transcriber test with an injected extractor) covering
+  incremental emission order, exact-start mapping with a `startSec` offset, and
+  ctx-cancel mid-pass.
+- Transcriber test asserting `Incremental` yields playhead-first chunk order
+  across the two passes and returns the complete cue set for transcript storage.
+- Confirm first `onChunk` fires after ~one chunk of audio, not after full-file
+  extraction (timing/ordering assertion via a fake extractor, not wall-clock).
+
+---
+
+## Multi-repo / client impact
+
+None expected to the contract. Reuses the existing notifier protocol
+(`TranslationStarted` / `TranslationCues` / `TranslationCompleted` /
+`TranslationFailed`, `notifier.go`) under the same translated-track `trackKey`.
+Translated cues simply begin arriving far earlier, and `Done/Total` keeps the
+chunk-granular meaning plain `transcribe` already uses (1c). Worth a smoke test
+on one `silo-android` / `silo-apple` client to confirm the live track renders
+cues arriving before the transcript track completes, and that `Done/Total` is
+treated as a fraction.
+
+## Risks & tradeoffs
+
+- **Translation context at chunk boundaries (live only).** Per-chunk translation
+  resets the batch context-neighbour window at each ~30s boundary — accepted
+  quality tradeoff for live; background jobs are unchanged.
+- **More/smaller ASR requests for live jobs.** 30s chunks mean more requests;
+  per-chunk timeout (`chunkSeconds * asrChunkTimeoutFactor`) and per-spawn
+  overhead scale; the 15s floor bounds the worst case.
+- **Incremental extraction reliability (Phase 2).** Segment-list polling + the
+  `-ss` seek path need fixtures for boundary timing, the final-segment race, and
+  cancel/cleanup. Single ffmpeg pass per direction keeps cost low.
+- **Boundary word-clips.** Smaller live chunks add more fixed-length boundaries
+  where a straddling word can clip — already a documented v1 limitation; the
+  silence-aligned follow-up would address it.
+
+## Out of scope / follow-ups
+
+- **Pace to playback (issue item 3):** rolling window ahead of the playhead to
+  bound work; separate change once interleave + incremental extraction land.
+- **Cross-chunk translation context:** feed the previous chunk's tail source
+  cues as untranslated context into each chunk's translate call.
+- **Silence-aligned chunk boundaries:** reduce boundary word-clips that smaller
+  live chunks make more frequent.
+
+## v1 scope note
+
+`docs/architecture/v1-scope.md` is **NOT LOCKED**; the issue carries
+`v1-proposed`. This is a **bug fix** to an already-shipped capability (no new
+user-facing capability, endpoint, or response field — only an internal config
+key and corrected streaming behavior), so it proceeds under the "bug fixes
+proceed normally" clause. Link the PR to the AI-subtitle capability item per the
+v1 PR requirements.
+
+## Pre-push checklist
+
+- `make lint`
+- `cd web && pnpm run lint && pnpm run format:check` (run only if web touched —
+  none expected)
+- `make verify-local-paths`
+- Go tests for `internal/subtitles/ai` and `internal/playback` in a
+  libvips-capable container (a bare-host `go test ./...` silently skips CGO
+  packages).
@@ -2915,6 +2915,7 @@ func effectiveSubtitleAIConfig(cfg *config.Config) (subtitleai.Config, string) {
 		ASRModel:              cfg.AI.ASRModel,
 		BatchSize:             cfg.SubtitleAI.BatchSize,
 		ContextNeighbors:      cfg.SubtitleAI.ContextNeighbors,
+		LiveASRChunkSeconds:   cfg.SubtitleAI.LiveASRChunkSeconds,
 		TranscribeQuotaJobs:   cfg.SubtitleAI.TranscribeQuotaJobs,
 		TranscribeQuotaPeriod: cfg.SubtitleAI.TranscribeQuotaPeriod,
 	}, disabledGateway
@@ -272,9 +272,12 @@ type SubtitleAIConfig struct {
 	BatchSize         int  `yaml:"-"`
 	ContextNeighbors  int  `yaml:"-"`
 	// ASRChunkSeconds is the audio chunk length per transcription request
-	// (60..600). Shorter chunks bound Whisper timestamp drift on long files;
+	// (15..600). Shorter chunks bound Whisper timestamp drift on long files;
 	// longer chunks mean fewer requests and fewer boundary word-clips.
 	ASRChunkSeconds int `yaml:"-"`
+	// LiveASRChunkSeconds is the smaller per-request chunk length for
+	// session-attached live subtitle jobs (15..600).
+	LiveASRChunkSeconds int `yaml:"-"`
 	// TranscribeQuotaJobs caps how many transcription jobs each non-admin user
 	// may start per rolling quota period (0 = unlimited).
 	TranscribeQuotaJobs int `yaml:"-"`
@@ -474,6 +474,11 @@ func LoadFromDB(m map[string]string) (*Config, error) {
 		return nil, err
 	}
 	cfg.SubtitleAI.ASRChunkSeconds = subtitleAIChunkSeconds
+	subtitleAILiveChunkSeconds, err := intOr(m, "subtitle_ai.live_asr_chunk_seconds", 30)
+	if err != nil {
+		return nil, err
+	}
+	cfg.SubtitleAI.LiveASRChunkSeconds = subtitleAILiveChunkSeconds
 	// A bad quota row must not block startup: fall back to "no quota" /
 	// the daily window with a warning instead of failing the config load.
 	transcribeQuotaJobs, err := intOr(m, "subtitle_ai.transcribe_quota_jobs", 0)
@@ -6,6 +6,7 @@ import (
 	"encoding/csv"
 	"encoding/json"
 	"fmt"
+	"io"
 	"math"
 	"os"
 	"os/exec"
@@ -13,6 +14,7 @@ import (
 	"sort"
 	"strconv"
 	"strings"
+	"time"
 )

 // AudioChunk is one extracted piece of an audio track. Start is the chunk's
@@ -40,25 +42,10 @@ func ExtractAudioChunks(ctx context.Context, filePath string, audioTrackIndex in
 	}

 	listPath := filepath.Join(dir, "segments.csv")
-	args := []string{
-		"-i", filePath,
-		"-vn", "-sn", "-dn",
-		"-map", fmt.Sprintf("0:a:%d", audioTrackIndex),
-		"-ac", "1",
-		"-ar", "16000",
-		"-c:a", "pcm_s16le",
-		"-f", "segment",
-		"-segment_time", strconv.Itoa(chunkSeconds),
-		"-segment_list", listPath,
-		"-segment_list_type", "csv",
-		"-y", filepath.Join(dir, "chunk%05d.wav"),
-	}
+	args := audioChunkExtractionArgs(filePath, audioTrackIndex, listPath,
+		filepath.Join(dir, "chunk%05d.wav"), 0, chunkSeconds)

-	ffmpeg := "ffmpeg"
-	if ffmpegPath != "" {
-		ffmpeg = ffmpegPath
-	}
-	cmd := exec.CommandContext(ctx, ffmpeg, args...)
+	cmd := exec.CommandContext(ctx, audioExtractionFFmpegBinary(ffmpegPath), args...)
 	var stderr bytes.Buffer
 	cmd.Stderr = &stderr

@@ -96,6 +83,126 @@ func ExtractAudioChunks(ctx context.Context, filePath string, audioTrackIndex in
 	return chunks, nil
 }

+// ExtractAudioChunksFrom extracts one audio track starting at startSec and calls
+// onSegment as each WAV segment is closed by ffmpeg. onSegment is synchronous:
+// if it blocks on ASR/translation, ffmpeg may continue extracting later chunks
+// into dir, so callers should remove processed chunks promptly. The caller owns
+// dir and cleanup. Segment starts are absolute media positions.
+func ExtractAudioChunksFrom(
+	ctx context.Context,
+	filePath string,
+	audioTrackIndex int,
+	dir, ffmpegPath string,
+	startSec float64,
+	chunkSeconds int,
+	onSegment func(AudioChunk) error,
+) error {
+	if chunkSeconds <= 0 {
+		chunkSeconds = 600
+	}
+	if audioTrackIndex < 0 {
+		audioTrackIndex = 0
+	}
+	if startSec < 0 {
+		startSec = 0
+	}
+
+	listPath := filepath.Join(dir, "segments.csv")
+	args := audioChunkExtractionArgs(filePath, audioTrackIndex, listPath,
+		filepath.Join(dir, "chunk%05d.wav"), startSec, chunkSeconds)
+
+	cmdCtx, cancel := context.WithCancel(ctx)
+	defer cancel()
+
+	cmd := exec.CommandContext(cmdCtx, audioExtractionFFmpegBinary(ffmpegPath), args...)
+	var stderr bytes.Buffer
+	cmd.Stderr = &stderr
+
+	if err := cmd.Start(); err != nil {
+		return fmt.Errorf("start ffmpeg audio extraction: %w", err)
+	}
+
+	waitCh := make(chan error, 1)
+	go func() {
+		waitCh <- cmd.Wait()
+	}()
+
+	seen := map[string]bool{}
+	tailer := &segmentListTailer{}
+	emit := func(flush bool) error {
+		return emitNewSegmentListEntries(listPath, dir, startSec, seen, tailer, flush, onSegment)
+	}
+
+	ticker := time.NewTicker(100 * time.Millisecond)
+	defer ticker.Stop()
+
+	for {
+		select {
+		case waitErr := <-waitCh:
+			segmentErr := emit(true)
+			if segmentErr != nil {
+				return segmentErr
+			}
+			if waitErr != nil {
+				if ctx.Err() != nil {
+					return ctx.Err()
+				}
+				return fmt.Errorf("ffmpeg audio extraction failed: %w (stderr: %s)",
+					waitErr, truncateStderr(stderr.String()))
+			}
+			if len(seen) == 0 {
+				return fmt.Errorf("ffmpeg produced no audio chunks for track %d", audioTrackIndex)
+			}
+			return nil
+		case <-ticker.C:
+			if err := emit(false); err != nil {
+				cancel()
+				<-waitCh
+				return err
+			}
+		case <-ctx.Done():
+			cancel()
+			<-waitCh
+			return ctx.Err()
+		}
+	}
+}
+
+func audioChunkExtractionArgs(
+	filePath string,
+	audioTrackIndex int,
+	listPath string,
+	outPattern string,
+	startSec float64,
+	chunkSeconds int,
+) []string {
+	args := []string{}
+	if startSec > 0 {
+		args = append(args, "-ss", strconv.FormatFloat(startSec, 'f', 3, 64))
+	}
+	args = append(args,
+		"-i", filePath,
+		"-vn", "-sn", "-dn",
+		"-map", fmt.Sprintf("0:a:%d", audioTrackIndex),
+		"-ac", "1",
+		"-ar", "16000",
+		"-c:a", "pcm_s16le",
+		"-f", "segment",
+		"-segment_time", strconv.Itoa(chunkSeconds),
+		"-segment_list", listPath,
+		"-segment_list_type", "csv",
+		"-y", outPattern,
+	)
+	return args
+}
+
+func audioExtractionFFmpegBinary(ffmpegPath string) string {
+	if ffmpegPath != "" {
+		return ffmpegPath
+	}
+	return "ffmpeg"
+}
+
 // parseSegmentList reads ffmpeg's CSV segment list (filename,start,end per
 // line) into a filename → start map. Best effort: a missing or malformed list
 // yields an empty map and callers fall back to nominal chunk starts.
@@ -123,6 +230,116 @@ func parseSegmentList(listPath string) map[string]float64 {
 	return starts
 }

+func emitNewSegmentListEntries(
+	listPath, dir string,
+	startOffset float64,
+	seen map[string]bool,
+	tailer *segmentListTailer,
+	flush bool,
+	onSegment func(AudioChunk) error,
+) error {
+	entries := tailer.read(listPath, flush)
+	for _, entry := range entries {
+		key := filepath.Base(entry.name)
+		if seen[key] {
+			continue
+		}
+		seen[key] = true
+		path := entry.name
+		if !filepath.IsAbs(path) {
+			path = filepath.Join(dir, path)
+		}
+		if onSegment != nil {
+			if err := onSegment(AudioChunk{Path: path, Start: startOffset + entry.start}); err != nil {
+				return err
+			}
+		}
+	}
+	return nil
+}
+
+type segmentListEntry struct {
+	name  string
+	start float64
+}
+
+type segmentListTailer struct {
+	offset  int64
+	pending string
+}
+
+func (t *segmentListTailer) read(listPath string, flush bool) []segmentListEntry {
+	f, err := os.Open(listPath)
+	if err != nil {
+		return nil
+	}
+	defer f.Close()
+
+	info, err := f.Stat()
+	if err == nil && info.Size() < t.offset {
+		t.offset = 0
+		t.pending = ""
+	}
+	if _, err := f.Seek(t.offset, io.SeekStart); err != nil {
+		return nil
+	}
+	data, err := io.ReadAll(f)
+	if err != nil {
+		return nil
+	}
+	t.offset += int64(len(data))
+
+	text := t.pending + string(data)
+	if text == "" {
+		return nil
+	}
+	lineEnd := strings.LastIndexByte(text, '\n')
+	if lineEnd < 0 {
+		if !flush {
+			t.pending = text
+			return nil
+		}
+		t.pending = ""
+		return parseSegmentListEntries(strings.NewReader(text))
+	}
+
+	complete := text[:lineEnd+1]
+	t.pending = text[lineEnd+1:]
+	if flush && t.pending != "" {
+		complete += t.pending
+		t.pending = ""
+	}
+	return parseSegmentListEntries(strings.NewReader(complete))
+}
+
+func parseSegmentListEntries(r io.Reader) []segmentListEntry {
+	reader := csv.NewReader(r)
+	var out []segmentListEntry
+	for {
+		row, err := reader.Read()
+		if err == io.EOF {
+			break
+		}
+		if err != nil {
+			// Keep completed rows; malformed rows are ignored best-effort.
+			break
+		}
+		if len(row) < 2 {
+			continue
+		}
+		start, err := strconv.ParseFloat(strings.TrimSpace(row[1]), 64)
+		if err != nil || start < 0 {
+			continue
+		}
+		name := strings.TrimSpace(row[0])
+		if name == "" {
+			continue
+		}
+		out = append(out, segmentListEntry{name: name, start: start})
+	}
+	return out
+}
+
 // maxPlausibleAudioDelay caps the probed audio/container start delta; beyond
 // this it is far more likely a probing artifact than a real stream delay.
 const maxPlausibleAudioDelay = 30.0
@@ -12,11 +12,12 @@ type Config struct {
 	TranslateEnabled bool
 	// TranscribeEnabled gates Whisper ASR generation (transcribe /
 	// transcribe_translate jobs).
-	TranscribeEnabled bool
-	ChatModel         string // chat-completions model used for translation
-	ASRModel          string // audio-transcription model used for ASR
-	BatchSize         int    // cues per translation request
-	ContextNeighbors  int    // preceding source cues sent as untranslated context
+	TranscribeEnabled   bool
+	ChatModel           string // chat-completions model used for translation
+	ASRModel            string // audio-transcription model used for ASR
+	BatchSize           int    // cues per translation request
+	ContextNeighbors    int    // preceding source cues sent as untranslated context
+	LiveASRChunkSeconds int    // smaller ASR chunks for session-attached live jobs
 	// TranscribeQuotaJobs caps how many transcription jobs (transcribe /
 	// transcribe_translate) each non-admin user may start per rolling quota
 	// period; 0 means unlimited.
@@ -309,7 +309,7 @@ func (s *Service) run(ctx context.Context, job *Job) {
 		TargetLanguage: job.TargetLanguage,
 	}, func(batch []SubtitleCue, done, total int) {
 		// Map cue progress into the 5%..95% band; push the batch live.
-		_ = s.repo.UpdateProgress(ctx, job.ID, JobStatusRunning, 0.05+0.9*float64(done)/float64(total), "Translating")
+		_ = s.repo.UpdateProgress(ctx, job.ID, JobStatusRunning, progressInBand(0.05, 0.9, done, total), "Translating")
 		if streaming {
 			s.notifier.TranslationCues(ctx, job.SessionID, job.MediaFileID, job.ID, trackKey,
 				toStreamCues(batch), done, total)
@@ -325,34 +325,9 @@ func (s *Service) run(ctx context.Context, job *Job) {

 	// Persist in chronological order regardless of the playhead-first order used
 	// for translation/streaming.
-	sortCuesByStart(translated)
-	sub, err := s.store.StoreSubtitle(storeCtx, subtitles.StoreSubtitleRequest{
-		MediaFileID: job.MediaFileID,
-		UserID:      job.RequestedBy,
-		Provider:    providerTranslated,
-		Language:    job.TargetLanguage,
-		Format:      subtitles.FormatSRT,
-		ReleaseName: releaseName,
-		Data:        SerializeSRT(translated),
-	})
-	if err != nil {
-		s.finishWithError(ctx, job, fmt.Errorf("store translated subtitle: %w", err))
+	if !s.finishTranslatedTrack(ctx, storeCtx, job, translated, releaseName, trackKey, streaming) {
 		return
 	}
-
-	if err := s.repo.CompleteJob(storeCtx, job.ID, sub.ID); err != nil {
-		s.logger.Warn("failed to mark subtitle ai job complete", "job", job.ID, "error", err)
-	}
-
-	if s.notifier != nil {
-		if streaming {
-			s.notifier.TranslationCompleted(storeCtx, job.SessionID, job.MediaFileID, job.ID, trackKey,
-				sub.ID, job.TargetLanguage, releaseName)
-		}
-		// Broadcast to other sessions watching this file (a no-op for the
-		// streaming requester, which already tracks its live track).
-		s.notifier.SubtitleReady(storeCtx, job.MediaFileID, sub.ID, job.TargetLanguage, releaseName)
-	}
 }

 // runTranscribe executes transcribe / transcribe_translate jobs: extract the
@@ -412,17 +387,37 @@ func (s *Service) runTranscribe(ctx context.Context, job *Job) {
 		s.notifier.TranslationStarted(ctx, job.SessionID, job.MediaFileID, job.ID, trackKey, liveLang, liveLabel, 0)
 	}

-	// Only a plain transcribe streams transcript cues — the chained kind
-	// streams the translated cues from its translation stage instead.
+	cfg := s.config()
+
+	// Plain transcribe streams transcript cues. A live chained job translates
+	// each non-empty ASR chunk immediately and streams translated cues from the
+	// same callback, while background chained jobs keep the whole-file path.
 	streamTranscript := streaming && job.Kind == JobKindTranscribe
+	streamTranslated := streaming && job.Kind == JobKindTranscribeTranslate
+	liveChunkSeconds := 0
+	if streaming {
+		liveChunkSeconds = cfg.LiveASRChunkSeconds
+	}
+	liveTranslate := liveTranslateState{sourceLanguage: hint}
 	cues, detected, err := s.transcriber.Transcribe(ctx, TranscribeJobRequest{
 		FilePath:        file.FilePath,
 		AudioTrackIndex: audioIdx,
 		LanguageHint:    hint,
 		StartPosition:   job.StartPosition,
-	}, func(chunk []SubtitleCue, done, total int) {
+		DurationSeconds: float64(file.Duration),
+		ChunkSeconds:    liveChunkSeconds,
+		Incremental:     streaming,
+	}, func(chunk []SubtitleCue, chunkLanguage string, done, total int) {
+		if liveTranslate.sourceLanguage == "" && chunkLanguage != "" {
+			liveTranslate.sourceLanguage = chunkLanguage
+		}
+		if streamTranslated {
+			s.translateLiveChunk(ctx, job, trackKey, chunk, done, total, &liveTranslate)
+			return
+		}
+
 		// Transcription occupies the 5%..70% progress band (chunk granularity).
-		_ = s.repo.UpdateProgress(ctx, job.ID, JobStatusRunning, 0.05+0.65*float64(done)/float64(total), "Transcribing")
+		_ = s.repo.UpdateProgress(ctx, job.ID, JobStatusRunning, progressInBand(0.05, 0.65, done, total), "Transcribing")
 		if streamTranscript {
 			s.notifier.TranslationCues(ctx, job.SessionID, job.MediaFileID, job.ID, trackKey,
 				toStreamCues(chunk), done, total)
@@ -476,6 +471,16 @@ func (s *Service) runTranscribe(ctx context.Context, job *Job) {
 		return
 	}

+	if streamTranslated {
+		if liveTranslate.err != nil {
+			s.finishWithErrorNotify(ctx, job, liveTranslate.err, !liveTranslate.failedNotified)
+			return
+		}
+		translatedLabel := translatedReleaseName(language, job.TargetLanguage)
+		s.finishTranslatedTrack(ctx, storeCtx, job, liveTranslate.cues, translatedLabel, trackKey, true)
+		return
+	}
+
 	// transcribe_translate: run the transcript through the regular translator.
 	// Cues already arrive playhead-first from chunk ordering, so the viewer's
 	// region translates (and streams) first here too.
@@ -485,7 +490,7 @@ func (s *Service) runTranscribe(ctx context.Context, job *Job) {
 		TargetLanguage: job.TargetLanguage,
 	}, func(batch []SubtitleCue, done, total int) {
 		// Translation occupies the 70%..95% band.
-		_ = s.repo.UpdateProgress(ctx, job.ID, JobStatusRunning, 0.7+0.25*float64(done)/float64(total), "Translating")
+		_ = s.repo.UpdateProgress(ctx, job.ID, JobStatusRunning, progressInBand(0.7, 0.25, done, total), "Translating")
 		if streaming {
 			s.notifier.TranslationCues(ctx, job.SessionID, job.MediaFileID, job.ID, trackKey,
 				toStreamCues(batch), done, total)
@@ -497,31 +502,86 @@ func (s *Service) runTranscribe(ctx context.Context, job *Job) {
 	}

 	storeCtx = context.WithoutCancel(ctx)
-	sortCuesByStart(translated)
 	translatedLabel := translatedReleaseName(language, job.TargetLanguage)
+	s.finishTranslatedTrack(ctx, storeCtx, job, translated, translatedLabel, trackKey, streaming)
+}
+
+type liveTranslateState struct {
+	sourceLanguage string
+	err            error
+	failedNotified bool
+	cues           []SubtitleCue
+}
+
+func (s *Service) translateLiveChunk(
+	ctx context.Context,
+	job *Job,
+	trackKey string,
+	chunk []SubtitleCue,
+	done, total int,
+	state *liveTranslateState,
+) {
+	_ = s.repo.UpdateProgress(ctx, job.ID, JobStatusRunning,
+		progressInBand(0.05, 0.9, done, total), "Transcribing & translating")
+	if state.err != nil || len(chunk) == 0 {
+		return
+	}
+	translatedChunk, err := s.translator.Translate(ctx, TranslateRequest{
+		Cues:           chunk,
+		SourceLanguage: state.sourceLanguage,
+		TargetLanguage: job.TargetLanguage,
+	}, func(batch []SubtitleCue, _, _ int) {
+		s.notifier.TranslationCues(ctx, job.SessionID, job.MediaFileID, job.ID, trackKey,
+			toStreamCues(batch), done, total)
+	})
+	if err != nil {
+		state.err = err
+		if !state.failedNotified {
+			s.notifier.TranslationFailed(ctx, job.SessionID, job.MediaFileID, job.ID,
+				trackKey, llm.Truncate(err.Error(), 500))
+			state.failedNotified = true
+		}
+		return
+	}
+	state.cues = append(state.cues, translatedChunk...)
+}
+
+func (s *Service) finishTranslatedTrack(
+	ctx context.Context,
+	storeCtx context.Context,
+	job *Job,
+	cues []SubtitleCue,
+	releaseName string,
+	trackKey string,
+	streaming bool,
+) bool {
+	sortCuesByStart(cues)
 	sub, err := s.store.StoreSubtitle(storeCtx, subtitles.StoreSubtitleRequest{
 		MediaFileID: job.MediaFileID,
 		UserID:      job.RequestedBy,
 		Provider:    providerTranslated,
 		Language:    job.TargetLanguage,
 		Format:      subtitles.FormatSRT,
-		ReleaseName: translatedLabel,
-		Data:        SerializeSRT(translated),
+		ReleaseName: releaseName,
+		Data:        SerializeSRT(cues),
 	})
 	if err != nil {
 		s.finishWithError(ctx, job, fmt.Errorf("store translated subtitle: %w", err))
-		return
+		return false
 	}
+
 	if err := s.repo.CompleteJob(storeCtx, job.ID, sub.ID); err != nil {
 		s.logger.Warn("failed to mark subtitle ai job complete", "job", job.ID, "error", err)
 	}
+
 	if s.notifier != nil {
 		if streaming {
 			s.notifier.TranslationCompleted(storeCtx, job.SessionID, job.MediaFileID, job.ID, trackKey,
-				sub.ID, job.TargetLanguage, translatedLabel)
+				sub.ID, job.TargetLanguage, releaseName)
 		}
-		s.notifier.SubtitleReady(storeCtx, job.MediaFileID, sub.ID, job.TargetLanguage, translatedLabel)
+		s.notifier.SubtitleReady(storeCtx, job.MediaFileID, sub.ID, job.TargetLanguage, releaseName)
 	}
+	return true
 }

 // defaultAudioTrackIndex returns the index of the default-flagged audio
@@ -544,6 +604,10 @@ func transcribedReleaseName(lang string) string {
 }

 func (s *Service) finishWithError(ctx context.Context, job *Job, err error) {
+	s.finishWithErrorNotify(ctx, job, err, true)
+}
+
+func (s *Service) finishWithErrorNotify(ctx context.Context, job *Job, err error, notify bool) {
 	status := JobStatusFailed
 	msg := llm.Truncate(err.Error(), 500)
 	// Only a genuine cancellation (user cancel, or shutdown via cancel) becomes
@@ -555,7 +619,7 @@ func (s *Service) finishWithError(ctx context.Context, job *Job, err error) {
 	if dbErr := s.repo.FailJob(context.WithoutCancel(ctx), job.ID, status, msg); dbErr != nil {
 		s.logger.Warn("failed to record subtitle ai job failure", "job", job.ID, "error", dbErr)
 	}
-	if job.SessionID != "" && s.notifier != nil {
+	if notify && job.SessionID != "" && s.notifier != nil {
 		s.notifier.TranslationFailed(context.WithoutCancel(ctx), job.SessionID, job.MediaFileID, job.ID,
 			liveTrackKey(job.ID), msg)
 	}
@@ -564,6 +628,13 @@ func (s *Service) finishWithError(ctx context.Context, job *Job, err error) {
 	}
 }

+func progressInBand(start, width float64, done, total int) float64 {
+	if total <= 0 {
+		return start
+	}
+	return start + width*float64(done)/float64(total)
+}
+
 // loadSource resolves the combined player subtitle index to translatable cues,
 // mirroring the index space used by the playback subtitle endpoints
 // (external → embedded → downloaded). Embedded tracks are extracted to SRT via
@@ -3,12 +3,16 @@ package ai
 import (
 	"context"
 	"errors"
+	"fmt"
+	"strings"
 	"sync"
 	"testing"
 	"time"

 	"github.com/Silo-Server/silo-server/internal/ai/jobrunner"
 	"github.com/Silo-Server/silo-server/internal/models"
+	"github.com/Silo-Server/silo-server/internal/playback"
+	"github.com/Silo-Server/silo-server/internal/subtitles"
 )

 // recordingRepo is a JobRepository that records ResetStaleJobs calls and
@@ -21,6 +25,20 @@ type recordingRepo struct {
 	inserts    int
 	quotaUsed  int // returned by CountTranscribeJobsByUserSince
 	lastJob    Job
+	completed  []int
+	failures   []recordedFailure
+	progress   []recordedProgress
+}
+
+type recordedFailure struct {
+	id      int64
+	status  JobStatus
+	message string
+}
+
+type recordedProgress struct {
+	progress float64
+	message  string
 }

 // InsertJob mirrors the Postgres repo's atomic quota guard: a non-nil quota
@@ -47,12 +65,25 @@ func (r *recordingRepo) CountTranscribeJobsByUserSince(context.Context, int, tim
 	defer r.mu.Unlock()
 	return r.quotaUsed, nil
 }
-func (r *recordingRepo) UpdateProgress(context.Context, int64, JobStatus, float64, string) error {
+func (r *recordingRepo) UpdateProgress(_ context.Context, _ int64, _ JobStatus, progress float64, message string) error {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	r.progress = append(r.progress, recordedProgress{progress: progress, message: message})
 	return nil
 }
-func (r *recordingRepo) CompleteJob(context.Context, int64, int) error           { return nil }
-func (r *recordingRepo) FailJob(context.Context, int64, JobStatus, string) error { return nil }
-func (r *recordingRepo) Heartbeat(context.Context, int64) error                  { return nil }
+func (r *recordingRepo) CompleteJob(_ context.Context, _ int64, subtitleID int) error {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	r.completed = append(r.completed, subtitleID)
+	return nil
+}
+func (r *recordingRepo) FailJob(_ context.Context, id int64, status JobStatus, message string) error {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	r.failures = append(r.failures, recordedFailure{id: id, status: status, message: message})
+	return nil
+}
+func (r *recordingRepo) Heartbeat(context.Context, int64) error { return nil }

 func (r *recordingRepo) ResetStaleJobs(_ context.Context, before time.Time, _ string) (int64, error) {
 	r.mu.Lock()
@@ -73,7 +104,7 @@ func (r *recordingRepo) snapshot() (int, time.Time) {
 type stubTranscriber struct{}

 func (stubTranscriber) Transcribe(context.Context, TranscribeJobRequest,
-	func([]SubtitleCue, int, int)) ([]SubtitleCue, string, error) {
+	TranscribeChunkCallback) ([]SubtitleCue, string, error) {
 	return nil, "", nil
 }

@@ -221,3 +252,404 @@ func TestRecoverReapsStaleJobsImmediately(t *testing.T) {
 		t.Errorf("stale cutoff = %v, want ~%v (now-staleJobThreshold)", before, want)
 	}
 }
+
+type runTranscribeFileResolver struct {
+	file *models.MediaFile
+}
+
+func (r runTranscribeFileResolver) GetByID(context.Context, int) (*models.MediaFile, error) {
+	return r.file, nil
+}
+
+type chunkedTranscriber struct {
+	chunks   [][]SubtitleCue
+	detected string
+	trace    *[]string
+	lastReq  TranscribeJobRequest
+}
+
+func (t *chunkedTranscriber) Transcribe(_ context.Context, req TranscribeJobRequest,
+	onChunk TranscribeChunkCallback) ([]SubtitleCue, string, error) {
+	t.lastReq = req
+	var all []SubtitleCue
+	for i, chunk := range t.chunks {
+		appendTrace(t.trace, fmt.Sprintf("transcribe:%d", i))
+		if onChunk != nil {
+			total := len(t.chunks)
+			if req.Incremental {
+				total = progressTotal(i+1, estimatedChunkTotal(req.DurationSeconds, req.ChunkSeconds))
+			}
+			onChunk(chunk, t.detected, i+1, total)
+		}
+		all = append(all, chunk...)
+	}
+	return all, t.detected, nil
+}
+
+type recordingTranslator struct {
+	calls     []TranslateRequest
+	errOnCall int
+	trace     *[]string
+}
+
+func (t *recordingTranslator) Translate(_ context.Context, req TranslateRequest,
+	onBatch func([]SubtitleCue, int, int)) ([]SubtitleCue, error) {
+	t.calls = append(t.calls, TranslateRequest{
+		Cues:           cloneTestCues(req.Cues),
+		SourceLanguage: req.SourceLanguage,
+		TargetLanguage: req.TargetLanguage,
+	})
+	call := len(t.calls)
+	if t.errOnCall == call {
+		return nil, errors.New("translator chunk failed")
+	}
+
+	out := cloneTestCues(req.Cues)
+	for i := range out {
+		for j, line := range out[i].Lines {
+			out[i].Lines[j] = req.TargetLanguage + ":" + line
+		}
+	}
+	if len(out) > 0 {
+		appendTrace(t.trace, "translate:"+strings.Join(out[0].Lines, " "))
+	}
+	if onBatch != nil {
+		onBatch(out, len(out), len(out))
+	}
+	return out, nil
+}
+
+type recordingSubtitleStore struct {
+	stored []subtitles.StoreSubtitleRequest
+}
+
+func (s *recordingSubtitleStore) StoreSubtitle(_ context.Context, req subtitles.StoreSubtitleRequest) (*subtitles.DownloadedSubtitle, error) {
+	s.stored = append(s.stored, req)
+	return &subtitles.DownloadedSubtitle{
+		ID:          len(s.stored),
+		MediaFileID: req.MediaFileID,
+		Provider:    req.Provider,
+		Language:    req.Language,
+		Format:      req.Format,
+		ReleaseName: req.ReleaseName,
+	}, nil
+}
+
+func (s *recordingSubtitleStore) GetSubtitleContent(context.Context, int) (*subtitles.DownloadedSubtitle, []byte, error) {
+	return nil, nil, errors.New("not implemented")
+}
+
+type notifierEvent struct {
+	kind    string
+	cues    []playback.StreamCue
+	done    int
+	total   int
+	message string
+}
+
+type recordingNotifier struct {
+	events []notifierEvent
+	trace  *[]string
+}
+
+func (n *recordingNotifier) SubtitleReady(_ context.Context, _ int, subtitleID int, language, _ string) {
+	n.events = append(n.events, notifierEvent{kind: fmt.Sprintf("ready:%d:%s", subtitleID, language)})
+}
+
+func (n *recordingNotifier) TranslationStarted(context.Context, string, int, int64, string, string, string, int) {
+	n.events = append(n.events, notifierEvent{kind: "started"})
+}
+
+func (n *recordingNotifier) TranslationCues(_ context.Context, _ string, _ int, _ int64, _ string,
+	cues []playback.StreamCue, done, total int) {
+	n.events = append(n.events, notifierEvent{kind: "cues", cues: cues, done: done, total: total})
+	if len(cues) > 0 {
+		appendTrace(n.trace, "stream:"+cues[0].Text)
+	}
+}
+
+func (n *recordingNotifier) TranslationCompleted(context.Context, string, int, int64, string, int, string, string) {
+	n.events = append(n.events, notifierEvent{kind: "completed"})
+}
+
+func (n *recordingNotifier) TranslationFailed(_ context.Context, _ string, _ int, _ int64, _, message string) {
+	n.events = append(n.events, notifierEvent{kind: "failed", message: message})
+}
+
+func TestRunTranscribeTranslateStreamingInterleavesTranslationPerChunk(t *testing.T) {
+	var trace []string
+	transcriber := &chunkedTranscriber{
+		detected: "en",
+		trace:    &trace,
+		chunks: [][]SubtitleCue{
+			{testCue(60, "second")},
+			{},
+			{testCue(0, "first")},
+		},
+	}
+	translator := &recordingTranslator{trace: &trace}
+	store := &recordingSubtitleStore{}
+	notifier := &recordingNotifier{trace: &trace}
+	repo := &recordingRepo{}
+	svc := newRunTranscribeService(repo, translator, transcriber, store, notifier, Config{LiveASRChunkSeconds: 30})
+
+	svc.runTranscribe(context.Background(), &Job{
+		ID:             7,
+		MediaFileID:    10,
+		Kind:           JobKindTranscribeTranslate,
+		SourceIndex:    -1,
+		TargetLanguage: "es",
+		SessionID:      "session-1",
+		StartPosition:  60,
+	})
+
+	if !transcriber.lastReq.Incremental {
+		t.Fatalf("streaming request did not enable incremental extraction")
+	}
+	if transcriber.lastReq.ChunkSeconds != 30 {
+		t.Fatalf("chunk override = %d, want 30", transcriber.lastReq.ChunkSeconds)
+	}
+	if got := len(translator.calls); got != 2 {
+		t.Fatalf("Translate calls = %d, want one per non-empty chunk", got)
+	}
+	if strings.Join(translator.calls[0].Cues[0].Lines, " ") != "second" {
+		t.Errorf("first translated chunk = %q, want playhead chunk", translator.calls[0].Cues[0].Lines)
+	}
+	if streamIdx, lastChunkIdx := traceIndex(trace, "stream:es:second"), traceIndex(trace, "transcribe:2"); streamIdx < 0 || lastChunkIdx < 0 || streamIdx > lastChunkIdx {
+		t.Fatalf("first translated stream did not precede final ASR chunk: %v", trace)
+	}
+
+	cueEvents := cueNotifierEvents(notifier.events)
+	if got := len(cueEvents); got != 2 {
+		t.Fatalf("TranslationCues events = %d, want 2", got)
+	}
+	if cueEvents[0].done != 1 || cueEvents[0].total != 6 || cueEvents[1].done != 3 || cueEvents[1].total != 6 {
+		t.Fatalf("chunk progress = (%d/%d, %d/%d), want 1/6 then 3/6",
+			cueEvents[0].done, cueEvents[0].total, cueEvents[1].done, cueEvents[1].total)
+	}
+	if len(store.stored) != 2 {
+		t.Fatalf("stored subtitles = %d, want transcript and translation", len(store.stored))
+	}
+	if store.stored[0].Provider != providerTranscribed || store.stored[1].Provider != providerTranslated {
+		t.Fatalf("stored providers = %q, %q", store.stored[0].Provider, store.stored[1].Provider)
+	}
+	translated, err := ParseCues(store.stored[1].Data)
+	if err != nil {
+		t.Fatalf("parse stored translation: %v", err)
+	}
+	if len(translated) != 2 || strings.Join(translated[0].Lines, " ") != "es:first" ||
+		strings.Join(translated[1].Lines, " ") != "es:second" {
+		t.Fatalf("stored translation not sorted/complete: %#v", translated)
+	}
+	if got := fmt.Sprint(repo.completed); got != "[2]" {
+		t.Fatalf("completed subtitle IDs = %s, want [2]", got)
+	}
+}
+
+func TestRunTranscribeTranslateStreamingFailureStoresTranscript(t *testing.T) {
+	var trace []string
+	transcriber := &chunkedTranscriber{
+		detected: "en",
+		trace:    &trace,
+		chunks: [][]SubtitleCue{
+			{testCue(0, "one")},
+			{testCue(30, "two")},
+			{testCue(60, "three")},
+		},
+	}
+	translator := &recordingTranslator{errOnCall: 2, trace: &trace}
+	store := &recordingSubtitleStore{}
+	notifier := &recordingNotifier{trace: &trace}
+	repo := &recordingRepo{}
+	svc := newRunTranscribeService(repo, translator, transcriber, store, notifier, Config{LiveASRChunkSeconds: 30})
+
+	svc.runTranscribe(context.Background(), &Job{
+		ID:             8,
+		MediaFileID:    10,
+		Kind:           JobKindTranscribeTranslate,
+		SourceIndex:    -1,
+		TargetLanguage: "es",
+		SessionID:      "session-1",
+	})
+
+	if got := len(translator.calls); got != 2 {
+		t.Fatalf("Translate calls = %d, want calls stop after failure", got)
+	}
+	if traceIndex(trace, "transcribe:2") < 0 {
+		t.Fatalf("ASR did not continue through final chunk after translate failure: %v", trace)
+	}
+	if len(store.stored) != 1 || store.stored[0].Provider != providerTranscribed {
+		t.Fatalf("stored subtitles = %#v, want transcript only", store.stored)
+	}
+	if len(repo.failures) != 1 || repo.failures[0].status != JobStatusFailed {
+		t.Fatalf("failures = %#v, want one failed job", repo.failures)
+	}
+	if got := countNotifierKind(notifier.events, "failed"); got != 1 {
+		t.Fatalf("TranslationFailed events = %d, want 1", got)
+	}
+	if len(repo.completed) != 0 {
+		t.Fatalf("completed despite translation failure: %v", repo.completed)
+	}
+}
+
+func TestRunTranscribeTranslateNonStreamingKeepsWholeFileTranslation(t *testing.T) {
+	transcriber := &chunkedTranscriber{
+		detected: "en",
+		chunks: [][]SubtitleCue{
+			{testCue(30, "later")},
+			{testCue(0, "earlier")},
+		},
+	}
+	translator := &recordingTranslator{}
+	store := &recordingSubtitleStore{}
+	notifier := &recordingNotifier{}
+	repo := &recordingRepo{}
+	svc := newRunTranscribeService(repo, translator, transcriber, store, notifier, Config{LiveASRChunkSeconds: 30})
+
+	svc.runTranscribe(context.Background(), &Job{
+		ID:             9,
+		MediaFileID:    10,
+		Kind:           JobKindTranscribeTranslate,
+		SourceIndex:    -1,
+		TargetLanguage: "es",
+	})
+
+	if transcriber.lastReq.Incremental {
+		t.Fatalf("background request unexpectedly enabled incremental extraction")
+	}
+	if transcriber.lastReq.ChunkSeconds != 0 {
+		t.Fatalf("background chunk override = %d, want 0", transcriber.lastReq.ChunkSeconds)
+	}
+	if got := len(translator.calls); got != 1 {
+		t.Fatalf("Translate calls = %d, want one whole-file call", got)
+	}
+	if got := len(translator.calls[0].Cues); got != 2 {
+		t.Fatalf("whole-file call cues = %d, want 2", got)
+	}
+	if got := countNotifierKind(notifier.events, "cues"); got != 0 {
+		t.Fatalf("non-streaming TranslationCues events = %d, want 0", got)
+	}
+	if len(store.stored) != 2 || store.stored[1].Provider != providerTranslated {
+		t.Fatalf("stored subtitles = %#v, want transcript and translated track", store.stored)
+	}
+}
+
+func TestRunTranscribeTranslateStreamingUsesDetectedLanguageForLiveChunks(t *testing.T) {
+	transcriber := &chunkedTranscriber{
+		detected: "ja",
+		chunks: [][]SubtitleCue{
+			{testCue(0, "konnichiwa")},
+		},
+	}
+	translator := &recordingTranslator{}
+	store := &recordingSubtitleStore{}
+	notifier := &recordingNotifier{}
+	repo := &recordingRepo{}
+	file := &models.MediaFile{
+		ID:          10,
+		FilePath:    "/media/movie.mkv",
+		Duration:    180,
+		AudioTracks: []models.AudioTrack{{Default: true}},
+	}
+	svc := newRunTranscribeServiceWithFile(repo, translator, transcriber, store, notifier,
+		Config{LiveASRChunkSeconds: 30}, file)
+
+	svc.runTranscribe(context.Background(), &Job{
+		ID:             10,
+		MediaFileID:    10,
+		Kind:           JobKindTranscribeTranslate,
+		SourceIndex:    -1,
+		TargetLanguage: "es",
+		SessionID:      "session-1",
+	})
+
+	if got := len(translator.calls); got != 1 {
+		t.Fatalf("Translate calls = %d, want one live chunk call", got)
+	}
+	if got := translator.calls[0].SourceLanguage; got != "ja" {
+		t.Fatalf("live chunk source language = %q, want detected ja", got)
+	}
+}
+
+func newRunTranscribeService(
+	repo *recordingRepo,
+	translator Translator,
+	transcriber Transcriber,
+	store *recordingSubtitleStore,
+	notifier Notifier,
+	cfg Config,
+) *Service {
+	file := &models.MediaFile{
+		ID:          10,
+		FilePath:    "/media/movie.mkv",
+		Duration:    180,
+		AudioTracks: []models.AudioTrack{{Language: "eng", Default: true}},
+	}
+	return newRunTranscribeServiceWithFile(repo, translator, transcriber, store, notifier, cfg, file)
+}
+
+func newRunTranscribeServiceWithFile(
+	repo *recordingRepo,
+	translator Translator,
+	transcriber Transcriber,
+	store *recordingSubtitleStore,
+	notifier Notifier,
+	cfg Config,
+	file *models.MediaFile,
+) *Service {
+	return NewService(context.Background(), cfg, repo, translator, transcriber, store, nil,
+		runTranscribeFileResolver{file: file}, notifier, "", nil, nil)
+}
+
+func testCue(startSeconds float64, text string) SubtitleCue {
+	start := time.Duration(startSeconds * float64(time.Second))
+	return SubtitleCue{
+		Start: start,
+		End:   start + 2*time.Second,
+		Lines: []string{text},
+	}
+}
+
+func cloneTestCues(cues []SubtitleCue) []SubtitleCue {
+	out := make([]SubtitleCue, len(cues))
+	for i, cue := range cues {
+		out[i] = cue
+		out[i].Lines = append([]string(nil), cue.Lines...)
+	}
+	return out
+}
+
+func appendTrace(trace *[]string, value string) {
+	if trace != nil {
+		*trace = append(*trace, value)
+	}
+}
+
+func traceIndex(trace []string, value string) int {
+	for i, got := range trace {
+		if got == value {
+			return i
+		}
+	}
+	return -1
+}
+
+func cueNotifierEvents(events []notifierEvent) []notifierEvent {
+	var out []notifierEvent
+	for _, event := range events {
+		if event.kind == "cues" {
+			out = append(out, event)
+		}
+	}
+	return out
+}
+
+func countNotifierKind(events []notifierEvent, kind string) int {
+	count := 0
+	for _, event := range events {
+		if event.kind == kind {
+			count++
+		}
+	}
+	return count
+}
@@ -2,7 +2,9 @@ package ai

 import (
 	"context"
+	"errors"
 	"fmt"
+	"math"
 	"os"
 	"path/filepath"
 	"strings"
@@ -21,9 +23,11 @@ const (
 	// 10 minutes of 16 kHz mono WAV is ~19 MB — under typical 25 MB API
 	// limits. Shorter chunks bound Whisper's within-chunk timestamp drift at
 	// the cost of more requests and more boundary word-clips; the operator
-	// tunes this via subtitle_ai.asr_chunk_seconds.
+	// tunes this via subtitle_ai.asr_chunk_seconds. Live jobs may use smaller
+	// chunks through subtitle_ai.live_asr_chunk_seconds to reduce time to first
+	// cue; the 15s floor bounds request count and boundary word-clips.
 	defaultASRChunkSeconds = 600
-	minASRChunkSeconds     = 60
+	minASRChunkSeconds     = 15
 	// Per-chunk request timeout: 3× realtime accommodates local Whisper
 	// servers on modest hardware.
 	asrChunkTimeoutFactor = 3
@@ -46,6 +50,9 @@ type TranscribeJobRequest struct {
 	AudioTrackIndex int     // resolved 0-based audio stream index
 	LanguageHint    string  // ISO 639-1; "" lets the model detect
 	StartPosition   float64 // seconds; chunks are processed playhead-first
+	DurationSeconds float64 // optional media duration for incremental progress
+	ChunkSeconds    int     // optional per-job override; 0 uses configured default
+	Incremental     bool    // stream extraction chunks instead of extracting whole file first
 }

 // Transcriber converts an audio track into subtitle cues. The built-in
@@ -55,9 +62,14 @@ type TranscribeJobRequest struct {
 // callers can report progress and stream cues live.
 type Transcriber interface {
 	Transcribe(ctx context.Context, req TranscribeJobRequest,
-		onChunk func(cues []SubtitleCue, done, total int)) ([]SubtitleCue, string, error)
+		onChunk TranscribeChunkCallback) ([]SubtitleCue, string, error)
 }

+// TranscribeChunkCallback receives each chunk's cues, the detected language for
+// that chunk when the ASR endpoint reports one, and chunk progress. A total of
+// 0 means the total is unknown/indeterminate.
+type TranscribeChunkCallback func(cues []SubtitleCue, language string, done, total int)
+
 // audioTranscriber is the slice of llm.Client the transcriber needs.
 type audioTranscriber interface {
 	Transcribe(ctx context.Context, req llm.TranscribeRequest) (*llm.Transcription, error)
@@ -76,18 +88,20 @@ type WhisperTranscriber struct {
 	ffmpegPath   atomic.Pointer[string]
 	chunkSeconds atomic.Int32
 	// extract and probeOffset are playback helpers, injectable for tests.
-	extract     func(ctx context.Context, filePath string, audioTrackIndex int, dir, ffmpegPath string, chunkSeconds int) ([]playback.AudioChunk, error)
-	probeOffset func(ctx context.Context, filePath string, audioTrackIndex int, ffmpegPath string) float64
+	extract            func(ctx context.Context, filePath string, audioTrackIndex int, dir, ffmpegPath string, chunkSeconds int) ([]playback.AudioChunk, error)
+	incrementalExtract func(ctx context.Context, filePath string, audioTrackIndex int, dir, ffmpegPath string, startSec float64, chunkSeconds int, onSegment func(playback.AudioChunk) error) error
+	probeOffset        func(ctx context.Context, filePath string, audioTrackIndex int, ffmpegPath string) float64
 }

 // NewWhisperTranscriber builds a transcriber backed by the shared AI client.
-// chunkSeconds outside [minASRChunkSeconds, defaultASRChunkSeconds] falls
-// back to the default (longer chunks would exceed upload limits).
+// chunkSeconds outside [minASRChunkSeconds, defaultASRChunkSeconds] clamps
+// to the nearest bound (longer chunks would exceed upload limits).
 func NewWhisperTranscriber(client *llm.Client, ffmpegPath string, chunkSeconds int) *WhisperTranscriber {
 	t := &WhisperTranscriber{
-		client:      client,
-		extract:     playback.ExtractAudioChunks,
-		probeOffset: playback.ProbeAudioStartOffset,
+		client:             client,
+		extract:            playback.ExtractAudioChunks,
+		incrementalExtract: playback.ExtractAudioChunksFrom,
+		probeOffset:        playback.ProbeAudioStartOffset,
 	}
 	t.SetExtraction(ffmpegPath, chunkSeconds)
 	return t
@@ -95,13 +109,10 @@ func NewWhisperTranscriber(client *llm.Client, ffmpegPath string, chunkSeconds i

 // SetExtraction updates the ffmpeg path and ASR chunk duration used by jobs
 // started afterwards. Safe for concurrent use; out-of-range chunk durations
-// fall back to the default, matching construction.
+// clamp to the nearest supported bound, matching per-request overrides.
 func (t *WhisperTranscriber) SetExtraction(ffmpegPath string, chunkSeconds int) {
-	if chunkSeconds < minASRChunkSeconds || chunkSeconds > defaultASRChunkSeconds {
-		chunkSeconds = defaultASRChunkSeconds
-	}
 	t.ffmpegPath.Store(&ffmpegPath)
-	t.chunkSeconds.Store(int32(chunkSeconds))
+	t.chunkSeconds.Store(int32(clampASRChunkSeconds(chunkSeconds)))
 }

 // Transcribe implements Transcriber. The returned cues are NOT sorted (they
@@ -109,7 +120,7 @@ func (t *WhisperTranscriber) SetExtraction(ffmpegPath string, chunkSeconds int)
 // endpoint's report for the first processed chunk, normalized to an ISO code
 // where possible.
 func (t *WhisperTranscriber) Transcribe(ctx context.Context, req TranscribeJobRequest,
-	onChunk func(cues []SubtitleCue, done, total int)) ([]SubtitleCue, string, error) {
+	onChunk TranscribeChunkCallback) ([]SubtitleCue, string, error) {
 	dir, err := os.MkdirTemp("", "silo-asr-*")
 	if err != nil {
 		return nil, "", fmt.Errorf("create ASR temp dir: %w", err)
@@ -120,50 +131,41 @@ func (t *WhisperTranscriber) Transcribe(ctx context.Context, req TranscribeJobRe
 	// even if the config reloads mid-job.
 	ffmpegPath := *t.ffmpegPath.Load()
 	chunkSeconds := int(t.chunkSeconds.Load())
-
-	chunks, err := t.extract(ctx, req.FilePath, req.AudioTrackIndex, dir, ffmpegPath, chunkSeconds)
-	if err != nil {
-		return nil, "", err
+	if req.ChunkSeconds > 0 {
+		chunkSeconds = req.ChunkSeconds
 	}
+	chunkSeconds = clampASRChunkSeconds(chunkSeconds)

 	// Audio streams can start after the container's timeline origin (TS
 	// remuxes especially); Whisper times are relative to the first audio
 	// sample, so the delta is a constant sync error unless added back.
 	startOffset := t.probeOffset(ctx, req.FilePath, req.AudioTrackIndex, ffmpegPath)

+	if req.Incremental {
+		return t.transcribeIncremental(ctx, req, dir, ffmpegPath, chunkSeconds, startOffset, onChunk)
+	}
+
+	chunks, err := t.extract(ctx, req.FilePath, req.AudioTrackIndex, dir, ffmpegPath, chunkSeconds)
+	if err != nil {
+		return nil, "", err
+	}
+
 	order := chunkOrderForPosition(chunks, req.StartPosition)
 	timeout := time.Duration(chunkSeconds*asrChunkTimeoutFactor) * time.Second

 	var all []SubtitleCue
 	detected := ""
 	for done, idx := range order {
-		if err := ctx.Err(); err != nil {
-			return nil, "", err
-		}
-		data, err := os.ReadFile(chunks[idx].Path)
-		if err != nil {
-			return nil, "", fmt.Errorf("read audio chunk: %w", err)
-		}
-		tr, err := t.client.Transcribe(ctx, llm.TranscribeRequest{
-			Filename: filepath.Base(chunks[idx].Path),
-			Audio:    data,
-			Language: req.LanguageHint,
-			Timeout:  timeout,
-		})
+		cues, lang, err := t.transcribeChunk(ctx, chunks[idx], req.LanguageHint, timeout, startOffset)
 		if err != nil {
 			return nil, "", fmt.Errorf("transcribe chunk %d/%d: %w", idx+1, len(chunks), err)
 		}
-		// Each chunk is read exactly once; deleting it as we go caps disk usage
-		// at one extraction rather than extraction + retranscription leftovers.
-		_ = os.Remove(chunks[idx].Path)
-
 		if detected == "" {
-			detected = normalizeDetectedLanguage(tr.Language)
+			detected = lang
 		}
-		cues := cuesFromSegments(tr.Segments, chunks[idx].Start+startOffset)
 		all = append(all, cues...)
 		if onChunk != nil {
-			onChunk(cues, done+1, len(order))
+			onChunk(cues, lang, done+1, len(order))
 		}
 	}

@@ -173,6 +175,165 @@ func (t *WhisperTranscriber) Transcribe(ctx context.Context, req TranscribeJobRe
 	return all, detected, nil
 }

+var errStopIncrementalPass = errors.New("stop incremental ASR pass")
+
+func (t *WhisperTranscriber) transcribeIncremental(
+	ctx context.Context,
+	req TranscribeJobRequest,
+	dir, ffmpegPath string,
+	chunkSeconds int,
+	startOffset float64,
+	onChunk TranscribeChunkCallback,
+) ([]SubtitleCue, string, error) {
+	timeout := time.Duration(chunkSeconds*asrChunkTimeoutFactor) * time.Second
+	pivot := incrementalPivotStart(req.StartPosition, req.DurationSeconds, chunkSeconds)
+	total := estimatedChunkTotal(req.DurationSeconds, chunkSeconds)
+
+	var all []SubtitleCue
+	detected := ""
+	done := 0
+	process := func(chunk playback.AudioChunk, chunkOffset float64) error {
+		cues, lang, err := t.transcribeChunk(ctx, chunk, req.LanguageHint, timeout, chunkOffset)
+		if err != nil {
+			return fmt.Errorf("transcribe chunk at %.3fs: %w", chunk.Start, err)
+		}
+		if detected == "" {
+			detected = lang
+		}
+		all = append(all, cues...)
+		done++
+		if onChunk != nil {
+			onChunk(cues, lang, done, progressTotal(done, total))
+		}
+		return nil
+	}
+
+	firstPassDone := done
+	firstPassOffset := audioOffsetForIncrementalPass(pivot, startOffset)
+	if err := t.runIncrementalPass(ctx, req, dir, ffmpegPath, pivot, chunkSeconds, func(chunk playback.AudioChunk) error {
+		return process(chunk, firstPassOffset)
+	}); err != nil {
+		if ctx.Err() != nil || pivot == 0 || done != firstPassDone {
+			return nil, "", err
+		}
+	}
+	if pivot > 0 {
+		wrap := func(chunk playback.AudioChunk) error {
+			if chunk.Start >= pivot {
+				return errStopIncrementalPass
+			}
+			return process(chunk, startOffset)
+		}
+		if err := t.runIncrementalPass(ctx, req, dir, ffmpegPath, 0, chunkSeconds, wrap); err != nil {
+			if !errors.Is(err, errStopIncrementalPass) {
+				return nil, "", err
+			}
+		}
+	}
+
+	if len(all) == 0 {
+		return nil, detected, fmt.Errorf("no speech recognized in the audio track")
+	}
+	return all, detected, nil
+}
+
+func (t *WhisperTranscriber) runIncrementalPass(
+	ctx context.Context,
+	req TranscribeJobRequest,
+	dir, ffmpegPath string,
+	startSec float64,
+	chunkSeconds int,
+	onSegment func(playback.AudioChunk) error,
+) error {
+	passDir := filepath.Join(dir, fmt.Sprintf("from-%.3f", startSec))
+	if err := os.MkdirAll(passDir, 0o755); err != nil {
+		return fmt.Errorf("create ASR incremental temp dir: %w", err)
+	}
+	return t.incrementalExtract(ctx, req.FilePath, req.AudioTrackIndex, passDir, ffmpegPath, startSec, chunkSeconds, onSegment)
+}
+
+func (t *WhisperTranscriber) transcribeChunk(
+	ctx context.Context,
+	chunk playback.AudioChunk,
+	languageHint string,
+	timeout time.Duration,
+	startOffset float64,
+) ([]SubtitleCue, string, error) {
+	if err := ctx.Err(); err != nil {
+		return nil, "", err
+	}
+	data, err := os.ReadFile(chunk.Path)
+	if err != nil {
+		return nil, "", fmt.Errorf("read audio chunk: %w", err)
+	}
+	tr, err := t.client.Transcribe(ctx, llm.TranscribeRequest{
+		Filename: filepath.Base(chunk.Path),
+		Audio:    data,
+		Language: languageHint,
+		Timeout:  timeout,
+	})
+	if err != nil {
+		return nil, "", err
+	}
+	// Each chunk is read exactly once; deleting it as we go caps disk usage
+	// at one extraction rather than extraction + retranscription leftovers.
+	_ = os.Remove(chunk.Path)
+
+	return cuesFromSegments(tr.Segments, chunk.Start+startOffset), normalizeDetectedLanguage(tr.Language), nil
+}
+
+func clampASRChunkSeconds(chunkSeconds int) int {
+	switch {
+	case chunkSeconds < minASRChunkSeconds:
+		return minASRChunkSeconds
+	case chunkSeconds > defaultASRChunkSeconds:
+		return defaultASRChunkSeconds
+	default:
+		return chunkSeconds
+	}
+}
+
+func incrementalPivotStart(startSeconds, durationSeconds float64, chunkSeconds int) float64 {
+	if startSeconds <= 0 || chunkSeconds <= 0 {
+		return 0
+	}
+	if durationSeconds > 0 && startSeconds >= durationSeconds {
+		startSeconds = math.Max(0, durationSeconds-0.001)
+	}
+	pivot := math.Floor(startSeconds/float64(chunkSeconds)) * float64(chunkSeconds)
+	if pivot < 0 {
+		return 0
+	}
+	return pivot
+}
+
+func estimatedChunkTotal(durationSeconds float64, chunkSeconds int) int {
+	if durationSeconds <= 0 || chunkSeconds <= 0 {
+		return 0
+	}
+	return int(math.Ceil(durationSeconds / float64(chunkSeconds)))
+}
+
+func progressTotal(done, total int) int {
+	if total <= 0 {
+		return 0
+	}
+	if total < done {
+		return done
+	}
+	return total
+}
+
+func audioOffsetForIncrementalPass(passStartSec, audioStartOffset float64) float64 {
+	if passStartSec <= 0 {
+		return audioStartOffset
+	}
+	if audioStartOffset <= passStartSec {
+		return 0
+	}
+	return audioStartOffset - passStartSec
+}
+
 // chunkOrderForPosition orders chunk indexes so the chunk containing
 // startSeconds is processed first, then forward, then wrapping to the start —
 // the viewer's current region fills first, mirroring translation's
@@ -117,7 +117,7 @@ func TestTranscribeProcessesChunksPlayheadFirst(t *testing.T) {
 	var chunkOrder []string
 	cues, _, err := tr.Transcribe(context.Background(), TranscribeJobRequest{
 		FilePath: "/x.mkv", StartPosition: 1300, // inside chunk 2
-	}, func(chunk []SubtitleCue, done, total int) {
+	}, func(chunk []SubtitleCue, _ string, done, total int) {
 		if total != 3 {
 			t.Errorf("total = %d, want 3", total)
 		}
@@ -157,6 +157,67 @@ func TestTranscribePassesHintAndChunkSizedTimeout(t *testing.T) {
 	}
 }

+func TestSetExtractionClampsChunkSecondsToNearestBound(t *testing.T) {
+	tr := &WhisperTranscriber{}
+
+	tr.SetExtraction("", 1)
+	if got := int(tr.chunkSeconds.Load()); got != minASRChunkSeconds {
+		t.Fatalf("low clamp = %d, want %d", got, minASRChunkSeconds)
+	}
+
+	tr.SetExtraction("", defaultASRChunkSeconds+1)
+	if got := int(tr.chunkSeconds.Load()); got != defaultASRChunkSeconds {
+		t.Fatalf("high clamp = %d, want %d", got, defaultASRChunkSeconds)
+	}
+
+	tr.SetExtraction("", 45)
+	if got := int(tr.chunkSeconds.Load()); got != 45 {
+		t.Fatalf("in-range chunk seconds = %d, want 45", got)
+	}
+}
+
+func TestTranscribeUsesRequestChunkSecondsOverride(t *testing.T) {
+	var dir string
+	var extractedChunkSeconds int
+	client := &fakeASRClient{
+		language: "en",
+		perChunk: map[string][]llm.TranscriptionSegment{
+			"chunk00000.wav": {{Start: 0, End: 1, Text: "hi"}},
+		},
+	}
+	tr := &WhisperTranscriber{
+		client: client,
+		extract: func(_ context.Context, _ string, _ int, d, _ string, chunkSeconds int) ([]playback.AudioChunk, error) {
+			dir = d
+			extractedChunkSeconds = chunkSeconds
+			p := filepath.Join(d, "chunk00000.wav")
+			if err := os.WriteFile(p, []byte("RIFF"), 0o644); err != nil {
+				return nil, err
+			}
+			return []playback.AudioChunk{{Path: p, Start: 0}}, nil
+		},
+		probeOffset: func(context.Context, string, int, string) float64 { return 0 },
+	}
+	tr.SetExtraction("", 600)
+
+	_, _, err := tr.Transcribe(context.Background(), TranscribeJobRequest{
+		FilePath:     "/x.mkv",
+		ChunkSeconds: 10,
+	}, nil)
+	if err != nil {
+		t.Fatalf("Transcribe: %v", err)
+	}
+	if extractedChunkSeconds != minASRChunkSeconds {
+		t.Fatalf("extract chunk seconds = %d, want clamped override %d", extractedChunkSeconds, minASRChunkSeconds)
+	}
+	if want := time.Duration(minASRChunkSeconds*asrChunkTimeoutFactor) * time.Second; client.requests[0].Timeout != want {
+		t.Fatalf("timeout = %v, want %v", client.requests[0].Timeout, want)
+	}
+	if _, statErr := os.Stat(dir); !os.IsNotExist(statErr) {
+		t.Errorf("temp dir %s not cleaned up", dir)
+	}
+}
+
 func TestTranscribeAllSilentChunksFailsClearly(t *testing.T) {
 	var dir string
 	client := &fakeASRClient{language: "en", perChunk: map[string][]llm.TranscriptionSegment{}}
@@ -326,6 +387,215 @@ func TestChunkOrderForPosition(t *testing.T) {
 	}
 }

+func TestTranscribeIncrementalStreamsPlayheadFirstAcrossTwoPasses(t *testing.T) {
+	var trace []string
+	var starts []float64
+	client := &fakeASRClient{
+		language: "en",
+		perChunk: map[string][]llm.TranscriptionSegment{
+			"inc1200.wav": {{Start: 0, End: 1, Text: "pivot"}},
+			"inc1800.wav": {{Start: 0, End: 1, Text: "tail"}},
+			"inc0.wav":    {{Start: 0, End: 1, Text: "head"}},
+			"inc600.wav":  {{Start: 0, End: 1, Text: "middle"}},
+		},
+	}
+	tr := &WhisperTranscriber{
+		client: client,
+		incrementalExtract: func(_ context.Context, _ string, _ int, dir, _ string, startSec float64, _ int,
+			onSegment func(playback.AudioChunk) error) error {
+			starts = append(starts, startSec)
+			segmentStarts := []float64{1200, 1800}
+			if startSec == 0 {
+				segmentStarts = []float64{0, 600, 1200}
+			}
+			for _, segStart := range segmentStarts {
+				name := fmt.Sprintf("inc%.0f.wav", segStart)
+				p := filepath.Join(dir, name)
+				if err := os.WriteFile(p, []byte("RIFF"), 0o644); err != nil {
+					return err
+				}
+				trace = append(trace, fmt.Sprintf("segment:%.0f", segStart))
+				if err := onSegment(playback.AudioChunk{Path: p, Start: segStart}); err != nil {
+					return err
+				}
+				trace = append(trace, fmt.Sprintf("after:%.0f", segStart))
+			}
+			return nil
+		},
+		probeOffset: func(context.Context, string, int, string) float64 { return 0 },
+	}
+	tr.SetExtraction("", 600)
+
+	var progress []string
+	cues, _, err := tr.Transcribe(context.Background(), TranscribeJobRequest{
+		FilePath:        "/x.mkv",
+		StartPosition:   1300,
+		DurationSeconds: 2400,
+		Incremental:     true,
+	}, func(chunk []SubtitleCue, _ string, done, total int) {
+		progress = append(progress, fmt.Sprintf("%d/%d", done, total))
+		trace = append(trace, "chunk:"+strings.Join(chunk[0].Lines, " "))
+	})
+	if err != nil {
+		t.Fatalf("Transcribe: %v", err)
+	}
+	if got := fmt.Sprint(starts); got != "[1200 0]" {
+		t.Fatalf("incremental pass starts = %s, want [1200 0]", got)
+	}
+	var requestOrder []string
+	for _, req := range client.requests {
+		requestOrder = append(requestOrder, req.Filename)
+	}
+	if got := strings.Join(requestOrder, ","); got != "inc1200.wav,inc1800.wav,inc0.wav,inc600.wav" {
+		t.Fatalf("request order = %s", got)
+	}
+	if got := strings.Join(progress, ","); got != "1/4,2/4,3/4,4/4" {
+		t.Fatalf("progress = %s, want chunk totals", got)
+	}
+	if chunkIdx, nextSegmentIdx := traceIndexString(trace, "chunk:pivot"), traceIndexString(trace, "segment:1800"); chunkIdx < 0 || nextSegmentIdx < 0 || chunkIdx > nextSegmentIdx {
+		t.Fatalf("first onChunk did not fire before the next extracted segment: %v", trace)
+	}
+	if len(cues) != 4 {
+		t.Fatalf("returned cues = %d, want complete transcript", len(cues))
+	}
+}
+
+func TestTranscribeIncrementalFallsBackToStartWhenSeekedPassProducesNoAudio(t *testing.T) {
+	client := &fakeASRClient{
+		language: "en",
+		perChunk: map[string][]llm.TranscriptionSegment{
+			"head.wav": {{Start: 1, End: 2, Text: "head"}},
+		},
+	}
+	tr := &WhisperTranscriber{
+		client: client,
+		incrementalExtract: func(_ context.Context, _ string, _ int, dir, _ string, startSec float64, _ int,
+			onSegment func(playback.AudioChunk) error) error {
+			if startSec > 0 {
+				return fmt.Errorf("ffmpeg produced no audio chunks for track 0")
+			}
+			p := filepath.Join(dir, "head.wav")
+			if err := os.WriteFile(p, []byte("RIFF"), 0o644); err != nil {
+				return err
+			}
+			return onSegment(playback.AudioChunk{Path: p, Start: 0})
+		},
+		probeOffset: func(context.Context, string, int, string) float64 { return 0 },
+	}
+	tr.SetExtraction("", 600)
+
+	cues, _, err := tr.Transcribe(context.Background(), TranscribeJobRequest{
+		FilePath:        "/x.mkv",
+		StartPosition:   1300,
+		DurationSeconds: 1800,
+		Incremental:     true,
+	}, nil)
+	if err != nil {
+		t.Fatalf("Transcribe: %v", err)
+	}
+	if len(cues) != 1 || strings.Join(cues[0].Lines, " ") != "head" {
+		t.Fatalf("fallback cues = %#v, want head cue from start pass", cues)
+	}
+	if len(client.requests) != 1 || client.requests[0].Filename != "head.wav" {
+		t.Fatalf("ASR requests = %#v, want only fallback chunk", client.requests)
+	}
+}
+
+func TestTranscribeIncrementalDoesNotDoubleApplyAudioOffsetAfterSeek(t *testing.T) {
+	client := &fakeASRClient{
+		language: "en",
+		perChunk: map[string][]llm.TranscriptionSegment{
+			"seek.wav": {{Start: 1, End: 2, Text: "seek"}},
+			"head.wav": {{Start: 1, End: 2, Text: "head"}},
+		},
+	}
+	tr := &WhisperTranscriber{
+		client: client,
+		incrementalExtract: func(_ context.Context, _ string, _ int, dir, _ string, startSec float64, _ int,
+			onSegment func(playback.AudioChunk) error) error {
+			name := "seek.wav"
+			chunkStart := 600.0
+			if startSec == 0 {
+				name = "head.wav"
+				chunkStart = 0
+			}
+			p := filepath.Join(dir, name)
+			if err := os.WriteFile(p, []byte("RIFF"), 0o644); err != nil {
+				return err
+			}
+			if err := onSegment(playback.AudioChunk{Path: p, Start: chunkStart}); err != nil {
+				return err
+			}
+			if startSec == 0 {
+				stop := filepath.Join(dir, "seek.wav")
+				if err := os.WriteFile(stop, []byte("RIFF"), 0o644); err != nil {
+					return err
+				}
+				return onSegment(playback.AudioChunk{Path: stop, Start: 600})
+			}
+			return nil
+		},
+		probeOffset: func(context.Context, string, int, string) float64 { return 1.25 },
+	}
+	tr.SetExtraction("", 600)
+
+	cues, _, err := tr.Transcribe(context.Background(), TranscribeJobRequest{
+		FilePath:        "/x.mkv",
+		StartPosition:   600,
+		DurationSeconds: 1200,
+		Incremental:     true,
+	}, nil)
+	if err != nil {
+		t.Fatalf("Transcribe: %v", err)
+	}
+	if len(cues) != 2 {
+		t.Fatalf("cues = %d, want seek and head", len(cues))
+	}
+	if want := 601 * time.Second; cues[0].Start != want {
+		t.Fatalf("seek cue start = %v, want %v without re-adding audio offset", cues[0].Start, want)
+	}
+	if want := time.Duration(2.25 * float64(time.Second)); cues[1].Start != want {
+		t.Fatalf("head cue start = %v, want %v with audio offset", cues[1].Start, want)
+	}
+}
+
+func TestTranscribeIncrementalKeepsProgressIndeterminateWhenDurationUnknown(t *testing.T) {
+	client := &fakeASRClient{
+		language: "en",
+		perChunk: map[string][]llm.TranscriptionSegment{
+			"head.wav": {{Start: 0, End: 1, Text: "head"}},
+		},
+	}
+	tr := &WhisperTranscriber{
+		client: client,
+		incrementalExtract: func(_ context.Context, _ string, _ int, dir, _ string, _ float64, _ int,
+			onSegment func(playback.AudioChunk) error) error {
+			p := filepath.Join(dir, "head.wav")
+			if err := os.WriteFile(p, []byte("RIFF"), 0o644); err != nil {
+				return err
+			}
+			return onSegment(playback.AudioChunk{Path: p, Start: 0})
+		},
+		probeOffset: func(context.Context, string, int, string) float64 { return 0 },
+	}
+	tr.SetExtraction("", 600)
+
+	var gotDone, gotTotal int
+	_, _, err := tr.Transcribe(context.Background(), TranscribeJobRequest{
+		FilePath:    "/x.mkv",
+		Incremental: true,
+	}, func(_ []SubtitleCue, _ string, done, total int) {
+		gotDone = done
+		gotTotal = total
+	})
+	if err != nil {
+		t.Fatalf("Transcribe: %v", err)
+	}
+	if gotDone != 1 || gotTotal != 0 {
+		t.Fatalf("progress = %d/%d, want 1/0 for indeterminate total", gotDone, gotTotal)
+	}
+}
+
 // Cue timing must use the muxer-reported chunk start (not index*chunkSeconds)
 // plus the probed audio start offset, or sync drifts on long files and
 // delayed-audio containers.
@@ -358,6 +628,15 @@ func TestTranscribeUsesExactChunkStartsAndAudioOffset(t *testing.T) {
 	}
 }

+func traceIndexString(values []string, target string) int {
+	for i, value := range values {
+		if value == target {
+			return i
+		}
+	}
+	return -1
+}
+
 func TestNormalizeDetectedLanguage(t *testing.T) {
 	cases := map[string]string{
 		"english":  "en",