Modelpublic

mja00/nemotron-speech-streaming-en-0.6b-webgpu

sourceHugging Faceotherupdated 5h agoView on Hugging Face

0likes

Model Card

Nemotron Speech Streaming EN 0.6B (560 ms), 4-bit ONNX for WebGPU

A browser-friendly build of nvidia/nemotron-speech-streaming-en-0.6b, a cache-aware FastConformer RNNT model with native streaming, punctuation, and capitalization. Used by Twitch Live Captions for on-device live captions.

It is derived from the sherpa-onnx export csukuangfj2/sherpa-onnx-nemotron-speech-streaming-en-0.6b-560ms-2026-04-25. The encoder's MatMul weights are quantized to 4 bits with ONNX Runtime's MatMulNBitsQuantizer (block size 32, symmetric), shrinking it from 2.4 GB to 680 MB. The decoder and joiner are unchanged fp32.

Earlier revisions of this repo (before this card mentioned 560 ms) contain the 1.12 s chunk build.

Files

File	Size	Role
`encoder.q4.onnx` + `encoder.q4.data`	680 MB	Cache-aware encoder, 0.56 s chunks. Run on WebGPU.
`decoder.onnx`	29 MB	RNNT prediction network (2-layer LSTM).
`joiner.onnx`	7 MB	RNNT joint network.
`tokens.txt`	9 KB	1024 SentencePiece tokens plus `<blk>` (id 1024).

Usage

—Features: 128-bin log-Mel, 16 kHz, 25 ms periodic Hann window, 10 ms hop, 512-point FFT, 0.97 pre-emphasis, Slaney mel scale and norm from 0 to 8000 Hz, log(max(power, FLT_EPSILON)), no normalization. This matches kaldi-native-fbank with sherpa-onnx's NeMo options.
—Encoder: audio_signal float32 [1, 128, 65], length int64 [1] = 65, plus the caches cache_last_channel [1, 24, 70, 1024], cache_last_time [1, 24, 1024, 8] and cache_last_channel_len int64 [1], all zero at start. Advance 56 feature frames per call and feed the *_next caches back in.
—Decoding: greedy RNNT with at most 10 symbols per encoder frame. Start the decoder with the blank token.

License

Licensed by NVIDIA Corporation under the NVIDIA Open Model License. See NOTICE.