Modelpublic

mja00/nemotron-3.5-asr-streaming-0.6b-webgpu

sourceHugging Faceotherupdated 3h agoView on Hugging Face

0likes

Model Card

Nemotron 3.5 ASR Streaming 0.6B (560 ms), 4-bit ONNX for WebGPU

A browser-friendly build of nvidia/nemotron-3.5-asr-streaming-0.6b, NVIDIA's multilingual cache-aware FastConformer RNNT with native streaming, punctuation, capitalization, and language-ID prompting (40 language-locales, 19 transcription-ready). Used by Twitch Live Captions for on-device live captions.

It is derived from the sherpa-onnx export csukuangfj2/sherpa-onnx-nemotron-3.5-asr-streaming-0.6b-560ms-2026-06-11. The encoder's MatMul weights are quantized to 4 bits with ONNX Runtime's MatMulNBitsQuantizer (block size 32, symmetric), shrinking it from 2.5 GB to 683 MB. The decoder and joiner are unchanged fp32.

Files

File	Size	Role
`encoder.q4.onnx` + `encoder.q4.data`	683 MB	Cache-aware encoder, 0.56 s chunks, with a `prompt_index` language input. Run on WebGPU.
`decoder.onnx`	60 MB	RNNT prediction network (2-layer LSTM).
`joiner.onnx`	38 MB	RNNT joint network.
`tokens.txt`	131 KB	13,087 tokens including `<xx-XX>` language tags, plus `<blk>` (id 13087).

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 (kaldi-native-fbank with sherpa-onnx's NeMo options).
—Encoder: audio_signal float32 [1, 128, 65], length int64 [1] = 65, the caches cache_last_channel [1, 24, 56, 1024], cache_last_time [1, 24, 1024, 8] and cache_last_channel_len int64 [1] (all zero at start), and prompt_index int64 [1]. Advance 56 feature frames per call and feed the *_next caches back in.
—Language: the prompt_dictionary in the encoder metadata maps locales to prompt ids, for example en-US = 0, es-ES = 2, ja-JP = 10, and auto = 101. In auto mode the model emits a <xx-XX> tag after terminal punctuation.
—Decoding: greedy RNNT with at most 10 symbols per encoder frame. Start the decoder with the blank token.

Auto-detection can misfire at the start of an utterance (for example, a few Korean characters on Japanese audio). Pass the locale explicitly when it is known.

License

Distributed under the OpenMDW License Agreement, version 1.1. See NOTICE for origin. NVIDIA's bias, explainability, privacy, and safety notes are included as published.