Fabio/voice_audio_multi
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions 1
voice_audio_multi/model_specs/lfm2_audio.json
Fabio c36510c21e
Some checks failed
Linux Build / Linux cpu (push) Has been cancelled
Details
Linux Build / Linux cpu-clang (push) Has been cancelled
Details
Linux Build / Linux vulkan (push) Has been cancelled
Details
mac-build / macOS CPU (push) Has been cancelled
Details
Nix Build / Nix (cpu) (push) Has been cancelled
Details
Nix Build / Nix (vulkan) (push) Has been cancelled
Details
windows-build / Windows CPU (push) Has been cancelled
Details
Initial import: audio.cpp with Orange Pi documentation
2026-11-04 19:05:22 +01:00

378 lines
13 KiB
JSON
Raw Permalink Blame History

{
"schema_version": 1,
"family": "lfm2_audio",
"display_name": "LFM2.5-Audio",
"description": "Liquid AI's LFM2.5-Audio, loaded from the llama.cpp-format GGUF components Liquid publishes. ASR: a FastConformer encoder feeds an LFM2 hybrid short-conv/attention backbone. TTS: the backbone generates audio frames through a depthformer, and an LFM2 detokenizer and ISTFT turn them into 24 kHz audio. S2S: a spoken question goes in as for ASR, and the reply comes out as interleaved text and audio.",
"category": "asr",
"status": "experimental",
"tasks": [
"asr",
"tts",
"s2s"
],
"modes": [
"offline",
"streaming"
],
"languages": [
"en",
"ja"
],
"capabilities": {
"asr": [
"vad_chunking"
],
"tts": [
"built_in_voices",
"long_form"
]
},
"options": {
"request": [
{
"name": "language",
"type": "string",
"description": "Must match the checkpoint's language (en or ja) when given; each published checkpoint transcribes and speaks one language.",
"required": false,
"default": "auto"
},
{
"name": "max_tokens",
"type": "int",
"description": "ASR: transcript tokens per audio chunk, 512 by default. TTS: audio frames (80 ms each) per text chunk, 512 by default. S2S: text tokens and audio frames of the reply together, 1024 by default. A transcript, a text chunk's speech or a reply that reaches it is cut off there, with a warning on stderr.",
"required": false,
"min": 1
},
{
"name": "audio_chunk_mode",
"type": "enum",
"description": "Audio chunking mode. auto transcribes input up to audio_chunk_duration_sec whole and splits longer input at pauses found by the bundled VAD (fixed chunks when the VAD model is absent); vad always splits at pauses and skips silence; fixed cuts every audio_chunk_duration_sec; none transcribes the whole input at once.",
"values": [
"auto",
"fixed",
"vad",
"none"
],
"required": false,
"default": "auto"
},
{
"name": "audio_chunk_duration_sec",
"type": "float",
"description": "Longest chunk in seconds, at least 1. With fixed chunks, a final piece shorter than 1 second and the chunk before it split their audio evenly.",
"required": false,
"min": 1,
"default": 30
},
{
"name": "temperature",
"type": "float",
"description": "TTS and S2S: audio code sampling temperature, 0.8 by default for TTS and 1.0 for S2S; 0 picks the most likely code.",
"required": false,
"min": 0
},
{
"name": "top_k",
"type": "int",
"description": "TTS and S2S: sample audio codes from the k most likely, 64 by default for TTS and 4 for S2S; 0 keeps all, 1 is greedy.",
"required": false,
"min": 0
},
{
"name": "text_temperature",
"type": "float",
"description": "S2S: text token sampling temperature; 0, the default, picks the most likely token, as liquid-audio's README and demo do.",
"required": false,
"min": 0,
"default": 0
},
{
"name": "text_top_k",
"type": "int",
"description": "S2S: with a text_temperature above 0, sample text tokens from the k most likely; 0, the default, keeps all, and 1 is greedy.",
"required": false,
"min": 0,
"default": 0
},
{
"name": "seed",
"type": "int",
"description": "TTS and S2S: sampling seed; omitted requests choose a random seed. TTS text chunk i uses seed + i. S2S text sampling draws from a stream of its own, so with the same seed a reply follows the greedy-text reply until a text token differs.",
"required": false,
"min": 0
},
{
"name": "text_chunk_mode",
"type": "enum",
"description": "TTS: long-form text chunking mode; japanese for the Japanese checkpoint and default otherwise when omitted.",
"values": [
"default",
"tag_aware",
"japanese",
"endline"
],
"required": false
},
{
"name": "text_chunk_size",
"type": "int",
"description": "TTS: maximum Unicode codepoints per text chunk, each spoken as its own turn.",
"required": false,
"min": 1,
"default": 200
},
{
"name": "stream_frames_per_event",
"type": "int",
"description": "TTS and S2S streaming: audio frames (80 ms each) per event; each event's audio is final. ASR runs offline only.",
"required": false,
"min": 1,
"default": 1
},
{
"name": "return_codes",
"type": "bool",
"description": "S2S: also return the reply's text tokens and audio codes as an lfm2_audio.reply artifact, to send back as history with the conversation's next turn.",
"required": false,
"default": false
}
],
"session": [
{
"name": "model_gguf",
"type": "string",
"description": "Backbone GGUF, relative to the model directory. Defaults to the directory's only backbone GGUF.",
"required": false,
"default": ""
},
{
"name": "mmproj_gguf",
"type": "string",
"description": "Audio encoder (mmproj) GGUF, relative to the model directory. Defaults to mmproj-<backbone GGUF name>, or else the directory's only mmproj GGUF.",
"required": false,
"default": ""
},
{
"name": "vocoder_gguf",
"type": "string",
"description": "TTS/S2S vocoder GGUF (depthformer), relative to the model directory. Defaults to vocoder-<backbone GGUF name>, or the directory's only vocoder GGUF.",
"required": false,
"default": ""
},
{
"name": "detokenizer_gguf",
"type": "string",
"description": "TTS/S2S audio detokenizer GGUF, relative to the model directory. Defaults to tokenizer-<backbone GGUF name>, or the directory's only tokenizer GGUF.",
"required": false,
"default": ""
},
{
"name": "vad_model_path",
"type": "string",
"description": "Path to the VAD model (any built-in VAD family; default Silero VAD) used by automatic audio chunking.",
"required": false,
"default": "assets/framework/models/silero_vad"
},
{
"name": "max_pass_duration_sec",
"type": "float",
"description": "Most audio, in seconds, the encoder takes in one pass: an ASR chunk (the whole input with audio_chunk_mode=none) or the S2S question. Longer audio is rejected before it is encoded.",
"required": false,
"min": 1,
"default": 120
},
{
"name": "cpu_repack",
"type": "bool",
"description": "On the CPU, multiply the matrices with the kernels of ggml's CPU extra buffer types where those take the weight's type and shape: the repacked ones take Q4_0 on x86 with AVX2 and Q4_0, Q8_0 and Q6_K on Arm; where ggml's AMX buffer is built, it comes first and also takes F16, Q8_0 and K-quants. Where the token embedding's type is taken, the text head gets a copy of it. Other backends ignore it.",
"required": false,
"default": true
},
{
"name": "conversation_cache_slots",
"type": "int",
"description": "S2S: how many conversations a session keeps the earlier questions' encoder output and prefill blocks of, for their next turns, the least recently used dropped first; 0 keeps none. Each kept conversation takes up to about 350 MiB of host memory at the conversation limit, and a slot that no conversation has used takes nothing. Output does not depend on it.",
"required": false,
"min": 0,
"default": 4
}
],
"load": []
},
"runtime": {
"tags": [
"gguf",
"cpu",
"metal",
"cuda"
]
},
"packages": [
{
"id": "lfm2_audio_1_5b_q8_0",
"display_name": "LFM2.5-Audio-1.5B Q8_0 GGUF",
"default": true,
"format": "gguf",
"precision": "q8_0",
"target_directory": "LFM2.5-Audio-1.5B-GGUF",
"files": [
"LFM2.5-Audio-1.5B-Q8_0.gguf",
"mmproj-LFM2.5-Audio-1.5B-Q8_0.gguf",
"vocoder-LFM2.5-Audio-1.5B-Q8_0.gguf",
"tokenizer-LFM2.5-Audio-1.5B-Q8_0.gguf"
],
"download": {
"kind": "huggingface_snapshot",
"repo": "LiquidAI/LFM2.5-Audio-1.5B-GGUF",
"revision": "7d525f883a077e20afb782f2ff618edcae0e39e4",
"gated": false
}
},
{
"id": "lfm2_audio_1_5b_f16",
"display_name": "LFM2.5-Audio-1.5B F16 GGUF",
"format": "gguf",
"precision": "f16",
"target_directory": "LFM2.5-Audio-1.5B-GGUF",
"files": [
"LFM2.5-Audio-1.5B-F16.gguf",
"mmproj-LFM2.5-Audio-1.5B-F16.gguf",
"vocoder-LFM2.5-Audio-1.5B-F16.gguf",
"tokenizer-LFM2.5-Audio-1.5B-F16.gguf"
],
"download": {
"kind": "huggingface_snapshot",
"repo": "LiquidAI/LFM2.5-Audio-1.5B-GGUF",
"revision": "7d525f883a077e20afb782f2ff618edcae0e39e4",
"gated": false
}
},
{
"id": "lfm2_audio_1_5b_q4_0",
"display_name": "LFM2.5-Audio-1.5B Q4_0 GGUF",
"format": "gguf",
"precision": "q4_0",
"target_directory": "LFM2.5-Audio-1.5B-GGUF",
"files": [
"LFM2.5-Audio-1.5B-Q4_0.gguf",
"mmproj-LFM2.5-Audio-1.5B-Q4_0.gguf",
"vocoder-LFM2.5-Audio-1.5B-Q4_0.gguf",
"tokenizer-LFM2.5-Audio-1.5B-Q4_0.gguf"
],
"download": {
"kind": "huggingface_snapshot",
"repo": "LiquidAI/LFM2.5-Audio-1.5B-GGUF",
"revision": "7d525f883a077e20afb782f2ff618edcae0e39e4",
"gated": false
}
},
{
"id": "lfm2_audio_1_5b_jp_q8_0",
"display_name": "LFM2.5-Audio-1.5B-JP Q8_0 GGUF",
"format": "gguf",
"precision": "q8_0",
"target_directory": "LFM2.5-Audio-1.5B-JP-GGUF",
"files": [
"LFM2.5-Audio-1.5B-JP-Q8_0.gguf",
"mmproj-LFM2.5-Audio-1.5B-JP-Q8_0.gguf",
"vocoder-LFM2.5-Audio-1.5B-JP-Q8_0.gguf",
"tokenizer-LFM2.5-Audio-1.5B-JP-Q8_0.gguf"
],
"download": {
"kind": "huggingface_snapshot",
"repo": "LiquidAI/LFM2.5-Audio-1.5B-JP-GGUF",
"revision": "64b96718b341dbd5650f9e85627cecdcbd4ac61b",
"gated": false
}
},
{
"id": "lfm2_audio_1_5b_jp_f16",
"display_name": "LFM2.5-Audio-1.5B-JP F16 GGUF",
"format": "gguf",
"precision": "f16",
"target_directory": "LFM2.5-Audio-1.5B-JP-GGUF",
"files": [
"LFM2.5-Audio-1.5B-JP-F16.gguf",
"mmproj-LFM2.5-Audio-1.5B-JP-F16.gguf",
"vocoder-LFM2.5-Audio-1.5B-JP-F16.gguf",
"tokenizer-LFM2.5-Audio-1.5B-JP-F16.gguf"
],
"download": {
"kind": "huggingface_snapshot",
"repo": "LiquidAI/LFM2.5-Audio-1.5B-JP-GGUF",
"revision": "64b96718b341dbd5650f9e85627cecdcbd4ac61b",
"gated": false
}
},
{
"id": "lfm2_audio_1_5b_jp_f32",
"display_name": "LFM2.5-Audio-1.5B-JP F32 GGUF",
"format": "gguf",
"precision": "f32",
"target_directory": "LFM2.5-Audio-1.5B-JP-GGUF",
"files": [
"LFM2.5-Audio-1.5B-JP-F32.gguf",
"mmproj-LFM2.5-Audio-1.5B-JP-F32.gguf",
"vocoder-LFM2.5-Audio-1.5B-JP-F32.gguf",
"tokenizer-LFM2.5-Audio-1.5B-JP-F32.gguf"
],
"download": {
"kind": "huggingface_snapshot",
"repo": "LiquidAI/LFM2.5-Audio-1.5B-JP-GGUF",
"revision": "64b96718b341dbd5650f9e85627cecdcbd4ac61b",
"gated": false
}
},
{
"id": "lfm2_audio_1_5b_jp_q4_0",
"display_name": "LFM2.5-Audio-1.5B-JP Q4_0 GGUF",
"format": "gguf",
"precision": "q4_0",
"target_directory": "LFM2.5-Audio-1.5B-JP-GGUF",
"files": [
"LFM2.5-Audio-1.5B-JP-Q4_0.gguf",
"mmproj-LFM2.5-Audio-1.5B-JP-Q4_0.gguf",
"vocoder-LFM2.5-Audio-1.5B-JP-Q4_0.gguf",
"tokenizer-LFM2.5-Audio-1.5B-JP-Q4_0.gguf"
],
"download": {
"kind": "huggingface_snapshot",
"repo": "LiquidAI/LFM2.5-Audio-1.5B-JP-GGUF",
"revision": "64b96718b341dbd5650f9e85627cecdcbd4ac61b",
"gated": false
}
}
],
"dependencies": [],
"ui": {
"tags": [
"ASR",
"TTS",
"GGUF",
"Stream"
],
"recommended_package": "lfm2_audio_1_5b_q8_0",
"docs": [
"docs/models/lfm2_audio.md"
],
"summary": "LFM2.5-Audio transcription, speech and spoken chat (English and Japanese checkpoints) from Liquid's published GGUFs.",
"builtin_voices": [
"us_male",
"us_female",
"uk_male",
"uk_female"
],
"default_voice": "us_male"
},
"sources": [
{
"format": "gguf",
"roots": {
"model": "."
},
"files": {},
"optional_files": {},
"tensors": {}
}
]
}
Reference in a new issue View git blame Copy permalink