Fabio
c36510c21e
Some checks failed
Linux Build / Linux cpu (push) Has been cancelled
Linux Build / Linux cpu-clang (push) Has been cancelled
Linux Build / Linux vulkan (push) Has been cancelled
mac-build / macOS CPU (push) Has been cancelled
Nix Build / Nix (cpu) (push) Has been cancelled
Nix Build / Nix (vulkan) (push) Has been cancelled
windows-build / Windows CPU (push) Has been cancelled
304 lines
8.3 KiB
JSON
304 lines
8.3 KiB
JSON
{
|
|
"schema_version": 1,
|
|
"family": "nemotron_asr",
|
|
"display_name": "Nemotron 3.5 ASR",
|
|
"description": "NVIDIA 600M streaming ASR model for low-latency and batch transcription across 40 language-locales, with native punctuation, capitalization, automatic language detection, and configurable chunk sizes.",
|
|
"category": "asr",
|
|
"status": "supported",
|
|
"tasks": [
|
|
"asr"
|
|
],
|
|
"modes": [
|
|
"offline",
|
|
"streaming"
|
|
],
|
|
"languages": [
|
|
"ar-AR",
|
|
"bg-BG",
|
|
"cs-CZ",
|
|
"da-DK",
|
|
"de-DE",
|
|
"el-GR",
|
|
"en-GB",
|
|
"en-US",
|
|
"es-ES",
|
|
"es-US",
|
|
"et-EE",
|
|
"fi-FI",
|
|
"fr-CA",
|
|
"fr-FR",
|
|
"he-IL",
|
|
"hi-IN",
|
|
"hr-HR",
|
|
"hu-HU",
|
|
"it-IT",
|
|
"ja-JP",
|
|
"ko-KR",
|
|
"lt-LT",
|
|
"lv-LV",
|
|
"mt-MT",
|
|
"nb-NO",
|
|
"nl-NL",
|
|
"nn-NO",
|
|
"pl-PL",
|
|
"pt-BR",
|
|
"pt-PT",
|
|
"ro-RO",
|
|
"ru-RU",
|
|
"sk-SK",
|
|
"sl-SI",
|
|
"sv-SE",
|
|
"th-TH",
|
|
"tr-TR",
|
|
"uk-UA",
|
|
"vi-VN",
|
|
"zh-CN"
|
|
],
|
|
"capabilities": {},
|
|
"dependencies": [],
|
|
"options": {
|
|
"request": [
|
|
{
|
|
"name": "language",
|
|
"type": "string",
|
|
"description": "ASR prompt language such as en-US, da-DK, or auto; default is the model's default prompt.",
|
|
"required": false
|
|
},
|
|
{
|
|
"name": "lookahead_tokens",
|
|
"type": "int",
|
|
"description": "Chunk-limited encoder right context in 80 ms frames; supported values and the default come from processor_config (0, 3, 6 or 13; default 3 for the published model). In masked speaker mode, omission derives lookahead from diarizer geometry (13 fallback).",
|
|
"required": false,
|
|
"min": 0
|
|
},
|
|
{
|
|
"name": "max_tokens",
|
|
"type": "int",
|
|
"description": "Maximum RNNT generated tokens; 0 or omitted uses the model-derived limit.",
|
|
"required": false,
|
|
"min": 0,
|
|
"default": 0
|
|
},
|
|
{
|
|
"name": "keep_language_tags",
|
|
"type": "bool",
|
|
"description": "Keep language tag tokens in decoded text; default false.",
|
|
"required": false,
|
|
"default": false
|
|
},
|
|
{
|
|
"name": "return_timestamps",
|
|
"type": "bool",
|
|
"description": "Compatibility/advisory flag; false does not suppress decoder token timestamps.",
|
|
"required": false,
|
|
"default": false
|
|
},
|
|
{
|
|
"name": "streaming",
|
|
"type": "bool",
|
|
"description": "Request a streaming run; must be false for offline sessions.",
|
|
"required": false,
|
|
"default": false
|
|
},
|
|
{
|
|
"name": "speaker_probabilities",
|
|
"type": "path",
|
|
"description": "nemotron_3_diar speaker_probabilities.safetensors for the same audio (diarize with return_frame_probabilities=true). Setting it enables speaker-tagged output: speaker turns with text and a SegLST artifact.",
|
|
"required": false
|
|
},
|
|
{
|
|
"name": "speaker_mode",
|
|
"type": "enum",
|
|
"values": [
|
|
"masked",
|
|
"attribution"
|
|
],
|
|
"description": "masked runs one ASR stream per active speaker on speaker-masked features (NeMo masked multitalker ASR) and transcribes overlapping speech; attribution assigns each word of one transcript to the most active speaker and is cheaper.",
|
|
"required": false,
|
|
"default": "masked"
|
|
},
|
|
{
|
|
"name": "speaker_mask",
|
|
"type": "enum",
|
|
"values": [
|
|
"mel",
|
|
"audio"
|
|
],
|
|
"description": "Masked mode only: mask log-mel features (NeMo mask_features) or zero the waveform of inactive 80 ms frames.",
|
|
"required": false,
|
|
"default": "mel"
|
|
},
|
|
{
|
|
"name": "speaker_segment_gap_sec",
|
|
"type": "float",
|
|
"description": "Start a new segment for a speaker after a pause longer than this; default 1.0 (NeMo sent_break_sec).",
|
|
"required": false,
|
|
"min": 0.0,
|
|
"default": 1.0
|
|
},
|
|
{
|
|
"name": "speaker_segment_max_sec",
|
|
"type": "float",
|
|
"description": "Streaming only: force-close a speaker-turn event after this many seconds; the final result keeps the full segments. Default 10. Must be positive.",
|
|
"required": false,
|
|
"default": 10
|
|
}
|
|
],
|
|
"session": [
|
|
{
|
|
"name": "weight_type",
|
|
"type": "enum",
|
|
"description": "Shared matmul weight storage type; default native.",
|
|
"preset": "weight_type_full",
|
|
"required": false,
|
|
"default": "native"
|
|
},
|
|
{
|
|
"name": "matmul_weight_type",
|
|
"type": "enum",
|
|
"description": "Encoder and decoder matmul weight storage type (native, f32, f16, bf16 or q8_0); defaults to weight_type.",
|
|
"preset": "weight_type_full",
|
|
"required": false
|
|
},
|
|
{
|
|
"name": "conv_weight_type",
|
|
"type": "enum",
|
|
"description": "Convolution weight storage type (native, f32 or f16); default native.",
|
|
"preset": "weight_type_conv",
|
|
"required": false,
|
|
"default": "native"
|
|
},
|
|
{
|
|
"name": "weight_context_mb",
|
|
"type": "int",
|
|
"description": "Weight context arena size in MiB; default 3072.",
|
|
"required": false,
|
|
"min": 1,
|
|
"default": 3072
|
|
},
|
|
{
|
|
"name": "encoder_graph_arena_mb",
|
|
"type": "int",
|
|
"description": "Encoder graph arena size in MiB; default 1024.",
|
|
"required": false,
|
|
"min": 1,
|
|
"default": 1024
|
|
},
|
|
{
|
|
"name": "decoder_graph_arena_mb",
|
|
"type": "int",
|
|
"description": "Decoder graph arena size in MiB; default 256.",
|
|
"required": false,
|
|
"min": 1,
|
|
"default": 256
|
|
},
|
|
{
|
|
"name": "mem_saver",
|
|
"type": "bool",
|
|
"description": "Release the offline encoder graph after each offline request; default false.",
|
|
"required": false,
|
|
"default": false
|
|
}
|
|
],
|
|
"load": []
|
|
},
|
|
"runtime": {
|
|
"tags": [
|
|
"gguf",
|
|
"stream"
|
|
]
|
|
},
|
|
"ui": {
|
|
"recommended_package": "nemotron_asr_q8_0",
|
|
"tags": [
|
|
"ASR",
|
|
"GGUF",
|
|
"Stream"
|
|
],
|
|
"docs": [
|
|
"docs/asr.md",
|
|
"docs/gguf.md"
|
|
]
|
|
},
|
|
"package_defaults": {
|
|
"download": {
|
|
"kind": "huggingface_snapshot",
|
|
"repo": "audio-cpp/audio.cpp-gguf",
|
|
"revision": "main",
|
|
"gated": false
|
|
}
|
|
},
|
|
"packages": [
|
|
{
|
|
"id": "nemotron_asr_q8_0",
|
|
"display_name": "Nemotron 3.5 ASR Streaming 0.6B Q8_0 GGUF",
|
|
"default": true,
|
|
"format": "gguf",
|
|
"precision": "q8_0",
|
|
"target_directory": "Nemotron-3.5-ASR-Streaming-0.6B-GGUF",
|
|
"files": [
|
|
"Nemotron-3.5-ASR-Streaming-0.6B-GGUF/nemotron-3.5-asr-streaming-0.6b-q8_0.gguf"
|
|
],
|
|
"strip_prefix": "Nemotron-3.5-ASR-Streaming-0.6B-GGUF"
|
|
},
|
|
{
|
|
"id": "nemotron_asr_f16",
|
|
"display_name": "Nemotron 3.5 ASR Streaming 0.6B F16 GGUF",
|
|
"format": "gguf",
|
|
"precision": "f16",
|
|
"target_directory": "Nemotron-3.5-ASR-Streaming-0.6B-GGUF",
|
|
"files": [
|
|
"Nemotron-3.5-ASR-Streaming-0.6B-GGUF/nemotron-3.5-asr-streaming-0.6b-f16.gguf"
|
|
],
|
|
"strip_prefix": "Nemotron-3.5-ASR-Streaming-0.6B-GGUF"
|
|
},
|
|
{
|
|
"id": "nemotron_asr_safetensors",
|
|
"display_name": "Nemotron 3.5 ASR Streaming 0.6B Safetensors",
|
|
"format": "safetensors",
|
|
"precision": "native",
|
|
"target_directory": "nemotron-3.5-asr-streaming-0.6b",
|
|
"files": [
|
|
"config.json",
|
|
"model.safetensors",
|
|
"processor_config.json",
|
|
"tokenizer.json"
|
|
],
|
|
"download": {
|
|
"kind": "huggingface_snapshot",
|
|
"repo": "nvidia/nemotron-3.5-asr-streaming-0.6b"
|
|
}
|
|
}
|
|
],
|
|
"sources": [
|
|
{
|
|
"format": "gguf",
|
|
"roots": {
|
|
"model": ".",
|
|
"weights": "$gguf"
|
|
},
|
|
"files": {
|
|
"config": "model:config.json",
|
|
"processor_config": "model:processor_config.json",
|
|
"tokenizer_json": "model:tokenizer.json"
|
|
},
|
|
"tensors": {
|
|
"weights": "weights:"
|
|
}
|
|
},
|
|
{
|
|
"format": "safetensors",
|
|
"roots": {
|
|
"model": "."
|
|
},
|
|
"files": {
|
|
"config": "model:config.json",
|
|
"processor_config": "model:processor_config.json",
|
|
"tokenizer_json": "model:tokenizer.json"
|
|
},
|
|
"tensors": {
|
|
"weights": "model:model.safetensors"
|
|
}
|
|
}
|
|
]
|
|
}
|