Nemotron 3.5 ASR Streaming 0.6B 窶・FP16 WebGPU ONNX
This repository is a custom three-session FP16 ONNX export of
nvidia/nemotron-3.5-asr-streaming-0.6b for browser WebGPU use.
- Source revision:
f3d333391852ba876df169dcc9ba902d25b6ab0b
- Default language used by the export probe:
ja-JP
- Encoder attention lookahead:
3 subsampled frames
- Weight/internal precision: FP16
- Graph I/O precision:
fp16
- Joint logits output:
fp32
Files
onnx/encoder_fp16.onnx: log-Mel features and language prompt -> projected encoder statesonnx/predictor_fp16.onnx: one RNNT predictor step with explicit LSTM hidden/cell stateonnx/joint_fp16.onnx: one RNNT joint step -> vocabulary logitsmanifest.json: exact source revision, graph contracts, package versions, and SHA-256 hashes
Large tensors are stored as ONNX external data. Keep every .onnx file together with its
referenced *_data file.
Important runtime note
Transformers.js does not currently include a native Nemotron3_5AsrForRNNT pipeline/model
implementation. This repository is therefore not loaded with a one-line
pipeline('automatic-speech-recognition', ...) call. Use the Transformers.js ONNX backend
(or ONNX Runtime Web directly) and implement the RNNT greedy loop across encoder, predictor,
and joint sessions.
The encoder graph is the cache-free, chunk-limited encoder path from the official Transformers
implementation. For continuous recognition, feed a rolling/padded log-Mel window and retain the
RNNT predictor state. It does not expose the official cache-aware encoder K/V and convolution
caches in this version of the package.
Input contract
Read manifest.json.contracts rather than hard-coding dimensions. The principal inputs are:
- encoder:
input_features, attention_mask, prompt_ids
- predictor:
input_ids, hidden_state, cell_state
- joint:
encoder_frame, predictor_output
The source processor uses 16 kHz audio and 128-bin log-Mel features. Copying the original
processor/tokenizer files into this repository does not by itself add a Nemotron processor class
to Transformers.js; the application must produce features matching processor_config.json.
License
The upstream model is distributed under Open Model Development and Weights License 1.1.
Review the upstream model card and license terms before redistribution or deployment.
