Dev/android to MAIN, last parakeet and diarziation - #19
Open
josep-bravo wants to merge 22 commits into
Open
josep-bravo wants to merge 22 commits into
josep-bravo wants to merge 22 commits into
Conversation
…mental decoding and enhance error logging
…_lock for thread safety
…trics and real-time factors
…and adjusting silence handling
…arization now correctly goes to false if stops few ms + best run on mp3 achieved
…edi that was non working DIART baseline still save, performs good on mp3 and direct mic audio, only thing wrong is between speakers random assignment at start, but performs very accurate separation of sentences and good assignemnt in general
- New asr_backend parameter (whisper | parakeet) in asr_params.yaml, overridable with ASR_BACKEND in Docker/.env through the asr_backend launch argument (docker-compose, android and mp3 compose files). - ParakeetASREngine subclasses ASREngine and only replaces model loading and the per-chunk transcription (_run_transcription hook). Buffering, VAD segmentation, speaker attribution, /speech_activity_detection and /speech_result behaviour are shared and unchanged. Whisper path unchanged. - Parakeet uses nvidia/parakeet-tdt-0.6b-v3, cached in speech_recognition/weights, float32, warm-up at load. Jetson Thor mp3 test: ~95 ms per chunk vs ~690 ms Whisper turbo. No Catalan support: keep asr_backend whisper where 'ca' matters. - ARM image: nemo_toolkit[asr]==2.4.0, huggingface_hub>=0.30,<1.0, pyarrow<20. Dockerfile.arm pins base torch/torchvision/numpy as pip constraints so no dependency can replace the NVIDIA builds. - docker-compose_mp3.yaml: default AUDIO_FILE now points to recordings/. - Unit tests for the Parakeet segment adapter, README docs. Based on Tomeu's Parakeet work in feature/parakeet-asr-backend. x86 requirements are not updated yet.
parakeet-tdt-0.6b-v3 neither outputs nor accepts a language id (confirmed by NVIDIA in NeMo issues #14799/#15097). Run NVIDIA langid_ambernet (107 languages incl. Catalan, ~6 ms per chunk) on each published chunk and pick the language from `language`, with the same meaning as the Whisper backend. Below parakeet_language_id_min_confidence (0.9) or for chunks under 0.25 s the last detected language is kept, so short or noisy replies do not flip it. - language_id.py: AmberNet wrapper, model cached in weights/nemo/ - New params parakeet_language_id_model, parakeet_language_id_min_confidence - Jetson Thor: FLEURS en/es/ca 88/96/99% at 1/2/3 s vs Whisper turbo 82/91/98% (Catalan 1 s: 88% vs 48%); ROS pipeline 46/46 on an es/ca/en sequence, ~97% on the English test mp3 - Unit tests for the language choice, README table
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
1 participant
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.
Joaan if you have gpu laptop downstairs, could you check if it builds and runs nicely as with jetson thor? :D