lucasw89/whisper_web
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
138 Commits 1 Branch 0 Tags
30969e86b71afcfd9b17e93dd6fbdf6fcbd299f9
Commit Graph
71 Commits
Author SHA1 Message Date
Lucas Winther 30969e86b7 Merge feat/parakeet-engine: Parakeet TDT 0.6B v3 English engine (spec 001 workstream 7) 
Co-Authored-By: Claude Opus 5.5 <[email protected]>

# Conflicts:
#	AGENTS.md
#	bun.lock
#	docs/Decisions.md
#	src/lib/models.ts
#	src/worker/inference.worker.ts
#	src/worker/protocol.ts
#	tests/models.test.ts
 2026-10-30 01:46:51 +01:00
Lucas Winther 9c89ab047f Merge feat/longform-pipeline: VAD long-form pipeline, opt-in (spec 001 workstream 3) 
Co-Authored-By: Claude Opus 5.5 <[email protected]>

# Conflicts:
#	docs/Architecture.md
#	docs/Decisions.md
#	src/App.tsx
#	src/index.css
#	src/lib/engine.ts
#	src/worker/inference.worker.ts
#	src/worker/protocol.ts
#	tests/engine.test.ts
#	tests/live-real-model.test.ts
#	tests/offline.md
#	tests/protocol.test.ts
 2026-10-30 01:39:17 +01:00
Lucas Winther 2c87471fdc Merge feat/vocabulary-hint: names-and-terms decoder prompt (spec 001 workstream 2) 
Co-Authored-By: Claude Opus 5.5 <[email protected]>

# Conflicts:
#	docs/Architecture.md
#	docs/Decisions.md
#	docs/WorkBreakdown.md
#	server-py/sidecar.py
#	src/App.tsx
#	src/server/live.ts
#	src/server/transcribe.ts
#	tests/fixtures/fake_sidecar.py
#	tests/transcribe.test.ts
 2026-10-30 01:35:28 +01:00
Lucas WintherandClaude Opus 5.5 5f335c3e41 Publish only whisper-small-danish under lucasw89; drop the turbo Danish entries 
The owner chose to publish just the small fine-tune: the turbo fine-tunes
showed no clear gain over stock turbo on da_festuge (D-WS6a).

Co-Authored-By: Claude Opus 5.5 <[email protected]>
 2026-10-30 01:32:30 +01:00
Lucas Winther aa7ae93c0c Merge feat/server-batched-vad: faster-whisper batched + VAD (spec 001 workstream 5) 
Co-Authored-By: Claude Opus 5.5 <[email protected]>

# Conflicts:
#	docs/Decisions.md
 2026-10-30 01:27:55 +01:00
Lucas Winther 439d099279 Merge feat/danish-models: Danish fine-tune catalog entries (spec 001 workstream 6) 
Co-Authored-By: Claude Opus 5.5 <[email protected]>

# Conflicts:
#	docs/Decisions.md
#	docs/WorkBreakdown.md
 2026-10-30 01:27:45 +01:00
Lucas Winther c94689e2e0 Merge feat/translate: canTranslate + Opus-MT transcript translation (spec 001 workstream 4) 
Co-Authored-By: Claude Opus 5.5 <[email protected]>

# Conflicts:
#	docs/Decisions.md
 2026-10-30 01:15:05 +01:00
Lucas WintherandClaude Opus 5.5 951aad8f55 Default the Danish conversions to q4 on WASM, as measured on the converted files 
Co-Authored-By: Claude Opus 5.5 <[email protected]>
 2026-10-30 01:12:28 +01:00
Lucas WintherandClaude Opus 5.5 cc8e031d6b Run WASM sessions at the default graph optimization level now that ort-web 1.31 loads the 8-bit decoders 
Co-Authored-By: Claude Opus 5.5 <[email protected]>
 2026-10-30 00:45:46 +01:00
Lucas WintherandClaude Opus 5.5 104c1c4e4d Record the batched-VAD measurements and check heartbeat, bursts and cancel against the real sidecar 
Co-Authored-By: Claude Opus 5.5 <[email protected]>
 2026-10-30 00:23:55 +01:00
Lucas WintherandClaude Opus 5.5 160298197e Reuse the language-detection encoder pass in long-form decodes 
Whisper's generate() keeps only forward_params from its kwargs, and that
list omits encoder_outputs, so the precomputed encoder output was dropped
and auto-detect cost three encoder passes for a one-chunk clip. The
real-model test now pins one encoder pass per decoded chunk.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
 2026-10-30 00:19:08 +01:00
Lucas WintherandClaude Opus 5.5 1c1806840e Tune the server engine's Silero VAD so batched file transcription stops dropping quiet speech 
Co-Authored-By: Claude Opus 5.5 <[email protected]>
 2026-10-30 00:07:05 +01:00
Lucas WintherandClaude Opus 5.5 fdc9b500c0 Check Parakeet chunks export to .srt and timestamped text, and record the headless Chrome run 
Co-Authored-By: Claude Opus 5.5 <[email protected]>
 2026-10-29 23:52:05 +01:00
Lucas WintherandClaude Opus 5.5 edb9b332b9 Cap the hint prompt at 120 tokens: a full 223-token prompt garbles small models 
Co-Authored-By: Claude Opus 5.5 <[email protected]>
 2026-10-29 23:51:47 +01:00
Lucas WintherandClaude Opus 5.5 dc15623f62 WIP: uncommitted work at pause (session limit) 
Co-Authored-By: Claude Opus 5.5 <[email protected]>
 2026-10-29 23:03:27 +01:00
Lucas WintherandClaude Opus 5.5 826bcd2e49 WIP: uncommitted work at pause (session limit) 
Co-Authored-By: Claude Opus 5.5 <[email protected]>
 2026-10-29 23:03:27 +01:00
Lucas WintherandClaude Opus 5.5 ec7ed36426 WIP: uncommitted work at pause (session limit) 
Co-Authored-By: Claude Opus 5.5 <[email protected]>
 2026-10-29 23:03:27 +01:00
Lucas WintherandClaude Opus 5.5 9f1d58c0b7 Keep max_new_tokens: 440 on Transformers.js 4.3.2 and pin the seek-loop condition it avoids 
Co-Authored-By: Claude Opus 5.5 <[email protected]>
 2026-10-29 22:59:35 +01:00
Lucas WintherandClaude Opus 5.5 8e0ecdaf85 Add the synthetic meeting fixture and long-form measurement script 
Co-Authored-By: Claude Opus 5.5 <[email protected]>
 2026-10-29 22:28:48 +01:00
Lucas WintherandClaude Opus 5.5 0684122f45 Document transcript translation: decision, architecture, manual offline check 
Co-Authored-By: Claude Opus 5.5 <[email protected]>
 2026-10-29 22:11:37 +01:00
Lucas WintherandClaude Opus 5.5 c110c9203f Add the Translate transcript panel with its own exports 
Co-Authored-By: Claude Opus 5.5 <[email protected]>
 2026-10-29 22:10:51 +01:00
Lucas WintherandClaude Opus 5.5 a0748b5fa6 Translate transcripts with Opus-MT in a dedicated translation worker 
Co-Authored-By: Claude Opus 5.5 <[email protected]>
 2026-10-29 22:10:51 +01:00
Lucas WintherandClaude Opus 5.5 708fae5778 Run Parakeet TDT in the inference worker via parakeet.js on the shared onnxruntime-web 
Adds parakeet.js 1.4.4 (MIT) as a devDependency bundled into the worker
only; a build plugin keeps its own onnxruntime-web copy out of the bundle.
Includes an opt-in real-model test (PARAKEET_REAL=1).

Co-Authored-By: Claude Opus 5.5 <[email protected]>
 2026-10-29 22:08:05 +01:00
Lucas WintherandClaude Opus 5.5 f55ea6ecf8 Run browser file transcription through the long-form pipeline when asked and the VAD model is available 
Adds an Advanced "Skip silence in uploaded files" option (off by default
pending measurement), the vad/longForm protocol fields, the old-path
fallback when the VAD model is missing, and a plain message for files
with no speech.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
 2026-10-29 22:02:47 +01:00
Lucas WintherandClaude Opus 5.5 0c3a83c23c Run Silero VAD and Whisper decodes for the long-form pipeline on Transformers.js 
Co-Authored-By: Claude Opus 5.5 <[email protected]>
 2026-10-29 22:02:47 +01:00
Lucas WintherandClaude Opus 5.5 2a4f5b16a5 Add Danish fine-tune catalog entries under a single pending Hugging Face owner constant 
Co-Authored-By: Claude Opus 5.5 <[email protected]>
 2026-10-29 22:00:33 +01:00
Lucas WintherandClaude Opus 5.5 f7d0525423 Wait for the PCM payload, not just the request, in the live WebSocket test (race under load) 
Co-Authored-By: Claude Opus 5.5 <[email protected]>
 2026-10-29 21:58:03 +01:00
Lucas WintherandClaude Opus 5.5 8236ea3504 End the hint prompt in a full stop: an unterminated list made Whisper drop punctuation and loop 
Co-Authored-By: Claude Opus 5.5 <[email protected]>
 2026-10-29 21:55:48 +01:00
Lucas WintherandClaude Opus 5.5 c5a57cd33a Keep Parakeet on WASM unless WebGPU is chosen explicitly, and never offer it on phones 
Co-Authored-By: Claude Opus 5.5 <[email protected]>
 2026-10-29 21:51:26 +01:00
Lucas WintherandClaude Opus 5.5 827e9a8dae Add a Cache-API model file downloader that reports load progress and caches only complete files 
Co-Authored-By: Claude Opus 5.5 <[email protected]>
 2026-10-29 21:50:40 +01:00
Lucas WintherandClaude Opus 5.5 c9fdab1726 Add Parakeet window planning, word-to-chunk mapping, vocab tokenizer and pinned asset URLs 
Co-Authored-By: Claude Opus 5.5 <[email protected]>
 2026-10-29 21:49:52 +01:00
Lucas WintherandClaude Opus 5.5 d554162a2c Add a model family field and the Parakeet v3 catalog entry with exact file sizes 
Co-Authored-By: Claude Opus 5.5 <[email protected]>
 2026-10-29 21:48:39 +01:00
Lucas WintherandClaude Opus 5.5 19318a56f2 Add the long-form transcription pipeline behind transcribeLongForm 
Co-Authored-By: Claude Opus 5.5 <[email protected]>
 2026-10-29 21:46:16 +01:00
Lucas WintherandClaude Opus 5.5 1c5952f212 Validate the vocabulary hint server-side and pass it to faster-whisper as initial_prompt 
Co-Authored-By: Claude Opus 5.5 <[email protected]>
 2026-10-29 21:45:39 +01:00
Lucas WintherandClaude Opus 5.5 ad2eb10a28 Add compression-ratio and log-prob quality checks for decode fallback 
Co-Authored-By: Claude Opus 5.5 <[email protected]>
 2026-10-29 21:44:41 +01:00
Lucas WintherandClaude Opus 5.5 221ac594eb Let the transcribe progress tracker follow variable-length windows 
Co-Authored-By: Claude Opus 5.5 <[email protected]>
 2026-10-29 21:44:19 +01:00
Lucas WintherandClaude Opus 5.5 753ea03979 Upgrade @huggingface/transformers to 4.3.2 and carry the sequential-session patch 
Co-Authored-By: Claude Opus 5.5 <[email protected]>
 2026-10-29 21:44:18 +01:00
Lucas WintherandClaude Opus 5.5 313892b6df Carry the vocabulary hint as prompt from the engines through the worker into every window 
Co-Authored-By: Claude Opus 5.5 <[email protected]>
 2026-10-29 21:44:06 +01:00
Lucas WintherandClaude Opus 5.5 87617bef4c Fix a type error in the hint truncation test 
Co-Authored-By: Claude Opus 5.5 <[email protected]>
 2026-10-29 21:43:19 +01:00
Lucas WintherandClaude Opus 5.5 e0a4b2796f Add pure VAD speech-timestamp and chunking functions for the long-form pipeline 
Port of faster-whisper 1.2.1 vad.py (checked identical on a random
20 000-window probability trace for both min-silence settings).

Co-Authored-By: Claude Opus 5.5 <[email protected]>
 2026-10-29 21:43:18 +01:00
Lucas WintherandClaude Opus 5.5 e6de097cfd Feed Whisper a decoder prompt through the ASR pipeline and strip it from the output 
Co-Authored-By: Claude Opus 5.5 <[email protected]>
 2026-10-29 21:43:11 +01:00
Lucas WintherandClaude Opus 5.5 942799d3cb Add the vocabulary hint's normalization, storage and prompt truncation rule 
Co-Authored-By: Claude Opus 5.5 <[email protected]>
 2026-10-29 21:42:25 +01:00
Lucas WintherandClaude Opus 5.5 82ed34c7f9 Plan transcript translation units and sentence pieces for Opus-MT 
Co-Authored-By: Claude Opus 5.5 <[email protected]>
 2026-10-29 21:39:37 +01:00
Lucas WintherandClaude Opus 5.5 02a5eab119 Offer Whisper's Translate to English only on models trained for it (canTranslate) 
Co-Authored-By: Claude Opus 5.5 <[email protected]>
 2026-10-29 21:37:28 +01:00
Lucas WintherandClaude Opus 5.5 afce794cbf Mark server-live sidecar passes as live mode and cover the stream heartbeat with a test 
Co-Authored-By: Claude Opus 5.5 <[email protected]>
 2026-10-29 21:37:01 +01:00
Lucas WintherandClaude Opus 5.5 dc64bb8198 Run server-engine file transcriptions through faster-whisper's batched pipeline with VAD 
Co-Authored-By: Claude Opus 5.5 <[email protected]>
 2026-10-29 21:36:21 +01:00
Lucas WintherandClaude Opus 5.5 4942461f79 Add typecheck + CI checks, light theme with self-hosted Inter, fix real-model test 
- Typecheck: typescript 7 devDependency, `bun run typecheck`; CI gains a
  `check` stage (install, typecheck, fast tests) gating image builds (D-038).
- Light theme: every colour is a token; prefers-color-scheme: light
  overrides them. Inter Variable self-hosted from @fontsource-variable/inter,
  kept out of Bun's CSS inlining and precached (scripts/fonts.ts, D-037).
  Backend select no longer overflows on narrow phones.
- tests/live-real-model.test.ts: time budget scales with clip length and
  en_alien gets a measured WER bar; it had failed since July (D-039).
- Delete the temporary isolate_wer2.ts; fix icon "not checked in" comments
  and the README's model-weights location.
- Agent docs: branch naming, typecheck, CI, Bun CSS gotcha, permissions.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
 2026-10-29 20:52:13 +01:00
Lucas Winther 500cf3b36e Don't show the live "recording"/listening indicator until the model is ready 
startLive()'s browser-engine branch flipped liveStatus to "listening" (the
red live-dot, "Listening..." hint) the instant the mic was granted, even if
the Whisper pipeline was still loading — the user saw an active-looking
session while nothing was being transcribed yet. transcribeLiveSegment()
already correctly awaits the model-ready gate before posting a segment, so
this was purely a misleading-status bug.

Adds a "loading-model" LiveStatus between "starting" and "listening";
startLive() now waits on the same browserModelReady() gate the transcribe
path uses before claiming "listening", with race-safe cleanup if the load
fails while waiting. Server-engine live is unaffected.

Adds Decisions.md D-034; updates tests/live.md's manual verification steps.
 2026-07-30 08:31:23 +02:00
Lucas Winther 7040f031a6 Usability and test infrastructure updates. 2026-07-29 09:20:43 +02:00
Lucas Winther 214a9ef714 Better error handling 2026-07-29 07:59:16 +02:00