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
318 lines
10 KiB
JSON
318 lines
10 KiB
JSON
{
|
|
"schema_version": 1,
|
|
"family": "f5_tts",
|
|
"display_name": "F5-TTS",
|
|
"description": "Community F5-TTS flow-matching diffusion transformer for zero-shot voice cloning, including Arabic via the Habibi-TTS finetune (SWivid). Text conv conditioner, RoPE DiT, CFM sampler, Vocos vocoder.",
|
|
"category": "tts",
|
|
"status": "community",
|
|
"tasks": [
|
|
"tts",
|
|
"clone"
|
|
],
|
|
"modes": [
|
|
"offline"
|
|
],
|
|
"languages": [
|
|
"en",
|
|
"ar"
|
|
],
|
|
"runtime": {
|
|
"tags": [
|
|
"server"
|
|
]
|
|
},
|
|
"capabilities": {
|
|
"clone": [
|
|
"speaker_reference"
|
|
]
|
|
},
|
|
"options": {
|
|
"request": [
|
|
{
|
|
"name": "reference_text",
|
|
"type": "string",
|
|
"description": "Transcript matching the reference voice audio; required by F5-TTS zero-shot cloning.",
|
|
"required": true
|
|
},
|
|
{
|
|
"name": "dialect",
|
|
"type": "string",
|
|
"description": "Habibi dialect token: UNK MSA SAU UAE ALG IRQ EGY MAR OMN TUN LEV SDN LBY; default UNK.",
|
|
"required": false,
|
|
"default": "UNK"
|
|
},
|
|
{
|
|
"name": "sway_sampling_coef",
|
|
"type": "float",
|
|
"description": "Sway sampling coefficient for inference timesteps; default -1.0. Advertised bounds are recommended settings, not parser-enforced limits.",
|
|
"required": false,
|
|
"min": -5,
|
|
"max": 5,
|
|
"default": -1
|
|
},
|
|
{
|
|
"name": "speed",
|
|
"type": "float",
|
|
"description": "Speech speed multiplier for duration estimation; default 1.0. Generated text chunks shorter than 10 characters use 0.3 instead. Advertised bounds are recommended settings, not parser-enforced limits.",
|
|
"required": false,
|
|
"min": 0.5,
|
|
"max": 2,
|
|
"default": 1
|
|
},
|
|
{
|
|
"name": "seed",
|
|
"type": "int",
|
|
"description": "Non-negative noise seed; omitted selects a random seed. Explicit zero is deterministic.",
|
|
"required": false,
|
|
"min": 0
|
|
},
|
|
{
|
|
"name": "num_inference_steps",
|
|
"type": "int",
|
|
"description": "CFM Euler steps (NFE); default 32. The parser rejects nonpositive counts.",
|
|
"required": false,
|
|
"min": 1,
|
|
"default": 32
|
|
},
|
|
{
|
|
"name": "guidance_scale",
|
|
"type": "float",
|
|
"description": "Classifier-free guidance strength; default 2.0. Advertised bounds are recommended settings, not parser-enforced limits.",
|
|
"required": false,
|
|
"min": 0,
|
|
"max": 10,
|
|
"default": 2
|
|
},
|
|
{
|
|
"name": "strip_diacritics",
|
|
"type": "bool",
|
|
"description": "Strip Arabic combining marks (harakat/tanwin/shadda/tatweel) before synthesis; Habibi is trained on undiacritized ASR transcripts and garbles diacritized input. Default true.",
|
|
"required": false,
|
|
"default": true
|
|
}
|
|
],
|
|
"session": [
|
|
{
|
|
"name": "lora",
|
|
"type": "path",
|
|
"description": "PEFT DiT LoRA adapter directory or checkpoint file; merged into session-owned weights and graphs. Vocos is unchanged.",
|
|
"required": false
|
|
},
|
|
{
|
|
"name": "lora_scale",
|
|
"type": "float",
|
|
"description": "Multiplier for the adapter's alpha/r scaling; default 1.",
|
|
"required": false,
|
|
"default": 1.0
|
|
},
|
|
{
|
|
"name": "vocos_path",
|
|
"type": "string",
|
|
"description": "Path to the Vocos vocoder checkpoint (vocos.safetensors); required unless placed next to the DiT checkpoint.",
|
|
"required": false
|
|
},
|
|
{
|
|
"name": "dialect",
|
|
"type": "string",
|
|
"description": "Default Habibi dialect token for requests that do not set it; default UNK.",
|
|
"required": false,
|
|
"default": "UNK"
|
|
},
|
|
{
|
|
"name": "frame_budget",
|
|
"type": "int",
|
|
"description": "Total mel frames per CFM pass (ref + generated); larger budgets allow sentence-scale chunks at higher VRAM. Must be in [256, 8192]; omitted uses 2048. Explicit zero is not accepted. On CPU the allocated DiT graph pool caps a pass at 960 frames (10.2 s), so larger values are clamped there; CUDA keeps the full range.",
|
|
"required": false,
|
|
"min": 256,
|
|
"max": 8192,
|
|
"default": 2048
|
|
}
|
|
],
|
|
"load": []
|
|
},
|
|
"package_defaults": {
|
|
"download": {
|
|
"kind": "huggingface_snapshot",
|
|
"repo": "trklou/audio.cpp",
|
|
"revision": "main",
|
|
"gated": false
|
|
}
|
|
},
|
|
"packages": [
|
|
{
|
|
"id": "habibi_unified",
|
|
"display_name": "Habibi-TTS Unified (Arabic, multi-dialect)",
|
|
"description": "Unified multi-dialect Arabic checkpoint from SWivid/Habibi-TTS, converted to GGUF (DiT transformer + Vocos vocoder namespaces, f32).",
|
|
"default": true,
|
|
"format": "gguf",
|
|
"precision": "orig",
|
|
"target_directory": "Habibi-TTS/Unified",
|
|
"files": [
|
|
"habibi-unified/habibi-unified-orig.gguf",
|
|
"habibi-unified/vocab.txt"
|
|
],
|
|
"strip_prefix": "habibi-unified"
|
|
},
|
|
{
|
|
"id": "vocos_mel_24khz",
|
|
"display_name": "Vocos mel 24kHz vocoder (GGUF)",
|
|
"description": "Standalone Vocos vocoder GGUF for use with the original safetensors checkpoints.",
|
|
"default": false,
|
|
"format": "gguf",
|
|
"precision": "orig",
|
|
"target_directory": "vocos-mel-24khz",
|
|
"files": [
|
|
"vocos-mel-24khz/vocos-mel-24khz-orig.gguf"
|
|
],
|
|
"strip_prefix": "vocos-mel-24khz"
|
|
},
|
|
{
|
|
"id": "habibi_alg",
|
|
"display_name": "Habibi-TTS ALG specialized checkpoint (GGUF)",
|
|
"description": "Single-dialect ALG checkpoint from SWivid/Habibi-TTS, converted to GGUF (DiT transformer + Vocos vocoder namespaces, f32). Stronger ALG accent than the unified model.",
|
|
"default": false,
|
|
"format": "gguf",
|
|
"precision": "orig",
|
|
"target_directory": "Habibi-TTS/Specialized/ALG",
|
|
"files": [
|
|
"habibi-alg/habibi-alg-orig.gguf",
|
|
"habibi-alg/vocab.txt"
|
|
],
|
|
"strip_prefix": "habibi-alg"
|
|
},
|
|
{
|
|
"id": "habibi_egy",
|
|
"display_name": "Habibi-TTS EGY specialized checkpoint (GGUF)",
|
|
"description": "Single-dialect EGY checkpoint from SWivid/Habibi-TTS, converted to GGUF (DiT transformer + Vocos vocoder namespaces, f32). Stronger EGY accent than the unified model.",
|
|
"default": false,
|
|
"format": "gguf",
|
|
"precision": "orig",
|
|
"target_directory": "Habibi-TTS/Specialized/EGY",
|
|
"files": [
|
|
"habibi-egy/habibi-egy-orig.gguf",
|
|
"habibi-egy/vocab.txt"
|
|
],
|
|
"strip_prefix": "habibi-egy"
|
|
},
|
|
{
|
|
"id": "habibi_irq",
|
|
"display_name": "Habibi-TTS IRQ specialized checkpoint (GGUF)",
|
|
"description": "Single-dialect IRQ checkpoint from SWivid/Habibi-TTS, converted to GGUF (DiT transformer + Vocos vocoder namespaces, f32). Stronger IRQ accent than the unified model.",
|
|
"default": false,
|
|
"format": "gguf",
|
|
"precision": "orig",
|
|
"target_directory": "Habibi-TTS/Specialized/IRQ",
|
|
"files": [
|
|
"habibi-irq/habibi-irq-orig.gguf",
|
|
"habibi-irq/vocab.txt"
|
|
],
|
|
"strip_prefix": "habibi-irq"
|
|
},
|
|
{
|
|
"id": "habibi_mar",
|
|
"display_name": "Habibi-TTS MAR specialized checkpoint (GGUF)",
|
|
"description": "Single-dialect MAR checkpoint from SWivid/Habibi-TTS, converted to GGUF (DiT transformer + Vocos vocoder namespaces, f32). Stronger MAR accent than the unified model.",
|
|
"default": false,
|
|
"format": "gguf",
|
|
"precision": "orig",
|
|
"target_directory": "Habibi-TTS/Specialized/MAR",
|
|
"files": [
|
|
"habibi-mar/habibi-mar-orig.gguf",
|
|
"habibi-mar/vocab.txt"
|
|
],
|
|
"strip_prefix": "habibi-mar"
|
|
},
|
|
{
|
|
"id": "habibi_msa",
|
|
"display_name": "Habibi-TTS MSA specialized checkpoint (GGUF)",
|
|
"description": "Single-dialect MSA checkpoint from SWivid/Habibi-TTS, converted to GGUF (DiT transformer + Vocos vocoder namespaces, f32). Stronger MSA accent than the unified model.",
|
|
"default": false,
|
|
"format": "gguf",
|
|
"precision": "orig",
|
|
"target_directory": "Habibi-TTS/Specialized/MSA",
|
|
"files": [
|
|
"habibi-msa/habibi-msa-orig.gguf",
|
|
"habibi-msa/vocab.txt"
|
|
],
|
|
"strip_prefix": "habibi-msa"
|
|
},
|
|
{
|
|
"id": "habibi_sau",
|
|
"display_name": "Habibi-TTS SAU specialized checkpoint (GGUF)",
|
|
"description": "Single-dialect SAU checkpoint from SWivid/Habibi-TTS, converted to GGUF (DiT transformer + Vocos vocoder namespaces, f32). Stronger SAU accent than the unified model.",
|
|
"default": false,
|
|
"format": "gguf",
|
|
"precision": "orig",
|
|
"target_directory": "Habibi-TTS/Specialized/SAU",
|
|
"files": [
|
|
"habibi-sau/habibi-sau-orig.gguf",
|
|
"habibi-sau/vocab.txt"
|
|
],
|
|
"strip_prefix": "habibi-sau"
|
|
},
|
|
{
|
|
"id": "habibi_uae",
|
|
"display_name": "Habibi-TTS UAE specialized checkpoint (GGUF)",
|
|
"description": "Single-dialect UAE checkpoint from SWivid/Habibi-TTS, converted to GGUF (DiT transformer + Vocos vocoder namespaces, f32). Stronger UAE accent than the unified model.",
|
|
"default": false,
|
|
"format": "gguf",
|
|
"precision": "orig",
|
|
"target_directory": "Habibi-TTS/Specialized/UAE",
|
|
"files": [
|
|
"habibi-uae/habibi-uae-orig.gguf",
|
|
"habibi-uae/vocab.txt"
|
|
],
|
|
"strip_prefix": "habibi-uae"
|
|
}
|
|
],
|
|
"dependencies": [],
|
|
"ui": {
|
|
"recommended_package": "habibi_unified",
|
|
"tags": [
|
|
"TTS",
|
|
"Clone"
|
|
],
|
|
"docs": [
|
|
"docs/models/f5_tts.md"
|
|
]
|
|
},
|
|
"sources": [
|
|
{
|
|
"format": "gguf",
|
|
"roots": {
|
|
"model": ".",
|
|
"weights": "$gguf"
|
|
},
|
|
"files": {
|
|
"vocab": "model:vocab.txt"
|
|
},
|
|
"tensors": {
|
|
"transformer": {
|
|
"source": "weights:",
|
|
"prefix": "transformer"
|
|
}
|
|
},
|
|
"optional_tensors": {
|
|
"vocos_vocoder": {
|
|
"source": "weights:",
|
|
"prefix": "vocos"
|
|
}
|
|
}
|
|
},
|
|
{
|
|
"format": "safetensors",
|
|
"roots": {
|
|
"model": "."
|
|
},
|
|
"files": {
|
|
"vocab": "model:vocab.txt"
|
|
},
|
|
"tensors": {
|
|
"transformer": {
|
|
"source": "model:model_200000.safetensors",
|
|
"prefix": "ema_model.transformer"
|
|
}
|
|
}
|
|
}
|
|
]
|
|
}
|