Nemotron 3.5 Asr Streaming GGUF

--- license: openmdw-1.1 language: - ar - bg - cs - da - de - el - en - es - et - fi - fr - he - hi - hr - hu - it - ja - ko - lt - lv - nb - nl - nn - pl - pt - ro - ru - sk - sl - sv - th - tr - uk - vi - zh tags: - asr - speech-recognition - gguf - streaming - fastconformer - rnnt - multilingual - crispasr pipeline_tag: automatic-speech-recognition base_model: nvidia/nemotron-3.5-asr-streaming-0.6b --- # Nemotron-3.5-ASR-Streaming-0.6B GGUF GGUF conversion of [nvidia/nemotron-3.5-asr-streaming-0.6b](https://huggingface.co/nvidia/nemotron-3.5-asr-streaming-0.6b) for use with [CrispASR](https://github.com/CrispStrobe/CrispASR). ## Model details **Architecture:** Cache-Aware Streaming FastConformer encoder (24 layers, d=1024, 8 heads) + RNN-T decoder (2-layer LSTM, hidden=640) + joint network (640 → 13088 vocab). **Languages:** 39 languages, selected via `prompt_kernel` MLP conditioning (one-hot lang → 2048-dim hidden → 1024-dim encoder conditioning): | Code | Language | Code | Language | Code | Language | |------|----------|------|----------|------|----------| | ar-AR | Arabic | fr-CA | French (CA) | nn-NO | Norwegian (NN) | | bg-BG | Bulgarian | fr-FR | French (FR) | pl-PL | Polish | | cs-CZ | Czech | he-IL | Hebrew | pt-BR | Portuguese (BR) | | da-DK | Danish | hi-IN | Hindi | pt-PT | Portuguese (PT) | | de-DE | German | hr-HR | Croatian | ro-RO | Romanian | | el-GR | Greek | hu-HU | Hungarian | ru-RU | Russian | | en-GB | English (GB) | it-IT | Italian | sk-SK | Slovak | | en-US | English (US) | ja-JP | Japanese | sl-SI | Slovenian | | es-ES | Spanish (ES) | ko-KR | Korean | sv-SE | Swedish | | es-US | Spanish (US) | lt-LT | Lithuanian | th-TH | Thai | | et-EE | Estonian | lv-LV | Latvian | tr-TR | Turkish | | fi-FI | Finnish | nb-NO | Norwegian (NB) | uk-UA | Ukrainian | | | | nl-NL | Dutch | vi-VN | Vietnamese | | | | | | zh-CN / zh-TW | Chinese | **Key properties:** - Sample rate: 16 kHz mono - 128 mel filterbank features, n_fft=512, hop=160 (10ms), win=400 (25ms) - 8x time downsampling (causal) → 80ms frame duration - Streaming: cache-aware attention with 4 context presets (see below) - Vocab: 13087 SentencePiece tokens + 1 blank (pure RNN-T, no TDT durations) - Native punctuation and capitalization - License: [OpenMDW-1.1](https://github.com/linux-foundation/open-model-developer-weight-license/blo...

Source context: 1720 downloads · 6 likes · Pipeline automatic-speech-recognition · Repo cstr/nemotron-3.5-asr-streaming-GGUF