Add ReazonSpeech as the Japanese transcription engine - #16
Open
tomoima525 wants to merge 2 commits into
Open
tomoima525 wants to merge 2 commits into
tomoima525 wants to merge 2 commits into
Conversation
When Language is set to Japanese, transcribe with ReazonSpeech (the k2 Zipformer from reazon-research/reazonspeech-k2-v2, INT8 ONNX via sherpa-onnx, CPU-only) instead of Whisper. On real Japanese speech it measures well under half whisper-turbo's character error rate (hayamimi's scorecard: CER 5.8% vs 13.8%) while decoding many times faster than realtime, which also collapses streaming mode's stop latency: the live segments are already decoded at full accuracy, so the refine pass short-circuits and only the LLM cleanup remains. - reazon: ReazonTranscriber, a duck-type drop-in for Transcriber (modified_beam_search; strips ReazonSpeech's TV-subtitle brackets; segments recordings over 30s with the energy VAD and joins the pieces) - models: download/extract the sherpa-onnx GitHub release tarball (~440MB) into Application Support, idempotent and thread-safe - pipeline: engine resolved per call so menu changes apply live; on load/download failure Whisper stands in (and refine resumes) until the model shows up cached - app: "Japanese Engine" menu; selecting ReazonSpeech (or switching Language to Japanese) prefetches the model with notifications; boot downloads it when configured and survives a failed fetch - cli: --engine on transcribe/stream for A/B testing, download --reazonspeech - ReazonSpeech emits no punctuation, so the cleanup LLM restores it (its gate always fires on unpunctuated text); documented in the README along with the mixed EN/JA caveat and the Whisper menu alternative Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01UitLA3APeum1DaiyajF46A
sherpa-onnx 1.13 split its native libraries (libonnxruntime and friends) into a separate sherpa-onnx-core package, and declares that dependency dynamically per platform -- uv's resolver misses it, so `uv sync` installed the Python binding without its dylibs and importing sherpa_onnx failed with "Library not loaded: @rpath/libonnxruntime.dylib". Declare sherpa-onnx-core explicitly so it lands in the lockfile. Also extract the ReazonSpeech tarball into a staging directory and rename it into place, so an interrupted extraction can never leave a truncated .onnx at the final path (which would fail every later load with an opaque onnxruntime/protobuf error until manually deleted). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01UitLA3APeum1DaiyajF46A
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What
When Language is set to Japanese, sotto now transcribes with ReazonSpeech (k2 Zipformer, INT8 ONNX via sherpa-onnx, Apache-2.0, CPU-only) instead of Whisper — the approach hayamimi uses, where it measures CER 5.8% vs whisper-turbo's 13.8% on real Japanese speech. Universal/English always use Whisper.
Design decisions (as discussed): default-on for ja sessions, ReazonSpeech end-to-end (drafts + finals), lazy ~440MB download from the sherpa-onnx GitHub release tarball, LLM cleaner for punctuation, exposed as a Japanese Engine menu.
How
reazon.py(new):ReazonTranscriber, a duck-type drop-in forTranscriber(warmup/transcribe).modified_beam_searchdecoding (hayamimi: CER 8.6→5.8 for +25% decode time); strips ReazonSpeech's TV-subtitle brackets; nono_speech_probexists so only the duration/RMS pre-gates apply (transducers return empty on silence rather than hallucinating). Recordings over 30s are cut at speech pauses by the existing energy VAD and decoded piecewise — offline zipformer decoding is meant for utterance-scale audio.models.py: download + extract the release tarball into~/Library/Application Support/sotto/models/, idempotent and lock-serialized (a menu prefetch can't race a dictation's lazy load), withtarfile'sdatafilter. Model files resolved by glob like hayamimi, since basenames vary between exports.pipeline.py: the engine is resolved per call, so Language/engine menu changes apply without a restart. If the model fails to download/load, Whisper stands in until the model shows up cached — and the streaming refine pass, which short-circuits when ReazonSpeech decodes the live segments (a second pass over the same audio with the same model adds nothing), resumes under the fallback. Net effect for ja streaming: stop latency is just the LLM pass.app.py: "Japanese Engine" submenu; picking ReazonSpeech or switching Language to Japanese prefetches the model in the background with notifications; boot downloads it when the config needs it and never bricks startup on a failed fetch.cli.py:sotto transcribe --engine reazonspeech|whisperand the same onstreamfor A/B testing on your own voice;sotto download --reazonspeech.Tradeoffs documented in the README
Testing
python3 -m py_compileon all modules;uv lockresolvessherpa-onnx 1.13.7(macOS arm64 wheels verified on PyPI for cp312).uv run sotto transcribe --language ja --engine reazonspeechvs--engine whisper(accuracy A/B),uv run sotto stream --language ja(draft cadence + no refine line), and a menu-driven download of the model.🤖 Generated with Claude Code
https://claude.ai/code/session_01UitLA3APeum1DaiyajF46A
Generated by Claude Code