Views
No views yet
nvidia/nemotron-3.5-asr-streaming-0.6b
(FastConformer encoder + RNN-Transducer decoder, ~600M parameters, 40
language-locales) for one-shot (whole-file) transcription on the CPU
execution provider.LICENSE and NOTICE (both MUST
travel with any redistribution).| File | Purpose |
|---|---|
manifest.json | Self-describing model manifest (schema v2) — the content-detection key and the tensor-I/O + feature + language binding |
encoder.onnx + encoder.onnx.data | FastConformer encoder with bundled language conditioning; weights in the external .data sidecar (loaded automatically by ONNX Runtime) |
decoder_joint.onnx | Fused RNNT prediction network + joint network (one greedy step) |
tokens.txt | SentencePiece id→piece map (one piece per line; the decoder's output id space) |
tokenizer.model | SentencePiece model (provenance) |
manifest.json, encoder.onnx,
encoder.onnx.data, decoder_joint.onnx, tokens.txt.manifest.json with
format == "onnx", schema_version == 2) and routes it to its ONNX engine. The
pipeline is: pure-Rust log-mel features (normalize: "NA", raw log-mel) → encoder
(x, x_len, prompt_index) → greedy RNNT loop over decoder_joint.manifest.json → language.map to the encoder's int64
prompt_index. A missing or unknown language surfaces InvalidArgument.
Automatic language detection (the model's prompt_index = 101) is intentionally
not exposed here.en-US, de-DE, fr-FR, ja-JP, plus bare
forms en, de, fr, …) across 40 language-locales. The full
key→id map is in manifest.json → language.map. Examples:en / en-US → 0 fr / fr-FR → 8 ru / ru-RU → 11
es-ES → 2 de / de-DE → 9 it / it-IT → 15
zh-CN → 4 ja / ja-JP → 10 ko / ko-KR → 14