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
271 lines
8.1 KiB
JSON
271 lines
8.1 KiB
JSON
{
|
|
"schema_version": 1,
|
|
"family": "haas_asr",
|
|
"display_name": "Haas-ja-v1.0",
|
|
"description": "Japanese FastConformer-TDT ASR with a Kimi Delta Attention predictor, using the shared Parakeet implementation.",
|
|
"category": "asr",
|
|
"status": "community",
|
|
"tasks": [
|
|
"asr"
|
|
],
|
|
"modes": [
|
|
"offline",
|
|
"streaming"
|
|
],
|
|
"languages": [
|
|
"ja"
|
|
],
|
|
"runtime": {
|
|
"tags": [
|
|
"gguf",
|
|
"stream"
|
|
]
|
|
},
|
|
"capabilities": {
|
|
"asr": [
|
|
"word_timestamps",
|
|
"partial_results",
|
|
"vad_chunking"
|
|
]
|
|
},
|
|
"options": {
|
|
"request": [
|
|
{
|
|
"name": "return_timestamps",
|
|
"type": "bool",
|
|
"description": "Accepted for CLI --words-out compatibility. Word timestamps are always returned, including when false or omitted.",
|
|
"required": false
|
|
},
|
|
{
|
|
"name": "max_tokens",
|
|
"type": "int",
|
|
"description": "Maximum TDT 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": "audio_chunk_mode",
|
|
"type": "enum",
|
|
"description": "Offline audio chunking mode. fixed selects overlapping long-form windows regardless of offline_mode; vad uses the configured VAD model (Silero by default) to skip silence; auto follows offline_mode. Offline input is downmixed and resampled as needed.",
|
|
"values": [
|
|
"auto",
|
|
"fixed",
|
|
"vad",
|
|
"none"
|
|
],
|
|
"required": false,
|
|
"default": "auto"
|
|
},
|
|
{
|
|
"name": "audio_chunk_duration_sec",
|
|
"type": "float",
|
|
"description": "VAD maximum chunk duration; omitted inherits the session duration. Fixed/long_form windows use the session setting, not this request override.",
|
|
"required": false,
|
|
"min": 0.001
|
|
}
|
|
],
|
|
"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; defaults to weight_type, which defaults to native. Q8_0 measured 1.79x faster on the tested CPU and changed roughly 8 percent of transcripts without moving aggregate word error rate.",
|
|
"preset": "weight_type_full",
|
|
"required": false
|
|
},
|
|
{
|
|
"name": "conv_weight_type",
|
|
"type": "enum",
|
|
"description": "Convolution weight storage type; default native.",
|
|
"preset": "weight_type_conv",
|
|
"required": false,
|
|
"default": "native"
|
|
},
|
|
{
|
|
"name": "perf_mode",
|
|
"type": "enum",
|
|
"description": "Encoder attention implementation. Default off uses the validated relative-attention path; flash_attention enables the fused implementation, which was numerically validated but slower on the tested hardware.",
|
|
"preset": "perf_mode_flash_attention",
|
|
"required": false,
|
|
"default": "off"
|
|
},
|
|
{
|
|
"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": "audio_chunk_duration_sec",
|
|
"type": "float",
|
|
"description": "Center-region duration for buffered streaming in seconds; default 2. Fixed context windows are re-encoded rather than cache-aware. Also sizes offline fixed/long_form windows.",
|
|
"required": false,
|
|
"min": 0.001,
|
|
"default": 2
|
|
},
|
|
{
|
|
"name": "left_context_sec",
|
|
"type": "float",
|
|
"description": "Past context included when re-encoding each buffered-streaming window in seconds; default 10. Also applies to offline fixed/long_form windows.",
|
|
"required": false,
|
|
"min": 0,
|
|
"default": 10
|
|
},
|
|
{
|
|
"name": "right_context_sec",
|
|
"type": "float",
|
|
"description": "Future lookahead included when re-encoding each buffered-streaming window in seconds; default 2 and adds equivalent partial-result latency. Also applies to offline fixed/long_form windows.",
|
|
"required": false,
|
|
"min": 0,
|
|
"default": 2
|
|
},
|
|
{
|
|
"name": "streaming_attention_mode",
|
|
"type": "enum",
|
|
"description": "Attention policy inside each buffered window. full_context preserves bidirectional attention over the bounded window.",
|
|
"values": [
|
|
"full_context"
|
|
],
|
|
"required": false,
|
|
"default": "full_context"
|
|
},
|
|
{
|
|
"name": "offline_mode",
|
|
"type": "enum",
|
|
"description": "Offline encoder scheduling. full_context encodes the whole utterance, long_form uses bounded overlapping windows, and auto selects long_form beyond audio_chunk_threshold_sec.",
|
|
"values": [
|
|
"full_context",
|
|
"long_form",
|
|
"auto"
|
|
],
|
|
"required": false,
|
|
"default": "full_context"
|
|
},
|
|
{
|
|
"name": "audio_chunk_threshold_sec",
|
|
"type": "float",
|
|
"description": "Duration threshold used by offline_mode=auto before switching to bounded-window long-form execution; default 30 seconds.",
|
|
"required": false,
|
|
"min": 0.001,
|
|
"default": 30.0
|
|
},
|
|
{
|
|
"name": "vad_model_path",
|
|
"type": "path",
|
|
"description": "VAD model path used by audio_chunk_mode=vad (any built-in VAD family); default assets/framework/models/silero_vad.",
|
|
"required": false,
|
|
"default": "assets/framework/models/silero_vad"
|
|
}
|
|
],
|
|
"load": []
|
|
},
|
|
"package_defaults": {
|
|
"download": {
|
|
"kind": "huggingface_snapshot",
|
|
"repo": "audio-cpp/audio.cpp-gguf",
|
|
"revision": "main",
|
|
"gated": false
|
|
}
|
|
},
|
|
"packages": [
|
|
{
|
|
"id": "haas_ja_v1_q8_0",
|
|
"display_name": "Haas-ja-v1.0 Q8_0 GGUF (Japanese KDA-TDT)",
|
|
"format": "gguf",
|
|
"precision": "q8_0",
|
|
"target_directory": "Haas-ja-v1.0-GGUF",
|
|
"files": [
|
|
"Haas-ja-v1.0-GGUF/haas-ja-v1.0-q8_0.gguf"
|
|
],
|
|
"strip_prefix": "Haas-ja-v1.0-GGUF",
|
|
"default": true
|
|
},
|
|
{
|
|
"id": "haas_ja_v1_f16",
|
|
"display_name": "Haas-ja-v1.0 F16 GGUF (Japanese KDA-TDT)",
|
|
"format": "gguf",
|
|
"precision": "f16",
|
|
"target_directory": "Haas-ja-v1.0-GGUF",
|
|
"files": [
|
|
"Haas-ja-v1.0-GGUF/haas-ja-v1.0-f16.gguf"
|
|
],
|
|
"strip_prefix": "Haas-ja-v1.0-GGUF",
|
|
"default": false
|
|
}
|
|
],
|
|
"dependencies": [],
|
|
"ui": {
|
|
"recommended_package": "haas_ja_v1_q8_0",
|
|
"tags": [
|
|
"ASR",
|
|
"GGUF"
|
|
],
|
|
"docs": [
|
|
"docs/models/haas.md"
|
|
]
|
|
},
|
|
"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"
|
|
}
|
|
}
|
|
]
|
|
}
|