VocalPy: a core Python package for acoustic comms research
summary
VocalPy packages domain types (Audio, Spectrogram, Annotation), wraps cetta, uses dask for parallel processing and exports datasets as a sqlite single-file database for sharing. Neural models moved from TensorFlow notebooks to PyTorch + Lightning (TweetingNet / RNNs and TCNs). Benchmarking shows a big drop in character error rate when increasing window 176 → 2000, and adding a loss term reduces over-segmentation; demos and reproducible pipelines are available in the repo and roadmap for feature expansion.
AI-written summary · about
chapters (8)
- 0:00VocalPy project intro
- 3:27Need for core Python package
- 6:16Tweeting neural network case
- 9:28Code challenges and fixes
- 13:00VocalPy library features
- 16:12Benchmarking neural models
- 19:49Songbird vocal learning measure
- 21:32Roadmap and community vision
related talks
25mBeyond Transcription: Building Voice AI That Understands Conversations
35mSpeech Synthesis with Tacotron2 and PyTorch
26mMake Python Talk, Make Python Listen
36mNVIDIA NeMo: Toolkit for Conversational AI
10mPyTorch at Dolby Labs
31m“Sorry, Could you repeat that again?” - Speech Recognition with Python
01h 10mFull Workshop: Realtime Voice AI
16mThe End of Awkward AI Transcriptions