Automatic Speech Recognition
NeMo
PyTorch
speaker-diarization
speaker-recognition
speech
audio
Transformer
FastConformer
Conformer
NEST
NeMo
Eval Results (legacy)
Instructions to use nvidia/diar_streaming_sortformer_4spk-v2 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- NeMo
How to use nvidia/diar_streaming_sortformer_4spk-v2 with NeMo:
import nemo.collections.asr as nemo_asr asr_model = nemo_asr.models.ASRModel.from_pretrained("nvidia/diar_streaming_sortformer_4spk-v2") transcriptions = asr_model.transcribe(["file.wav"]) - Notebooks
- Google Colab
- Kaggle
Model preprocessor that computes mel spectogram is not streaming
#12
by ashu18180 - opened
The MEL Spectrogram pre-processor for NeMo implementation is not streaming.
You can easily have Claude, Codex, or your favorite coding agent make a streaming version. In the prompt, just specify that the streaming version should not wait for more audio than necessary for the next chunk, and that it must match the offline version with a maximum absolute error of 1e-6. Or you can do it yourself (though, from experience implementing it in Swift, it's a bit of a headache to work out the indexing logic yourself).