vllm-omni/benchmarks/tts/model_configs.yaml

# Universal TTS benchmark model registry.
# Maps HuggingFace model ID → supported tasks + per-task extra body fields.
# To add a new TTS model: add an entry here. No code changes required.
#
# The server auto-loads its Deploy YAML from vllm_omni/deploy/<model>.yaml via
# the Pipeline + Deploy schema introduced in #2383, so no stage_config path
# is tracked here.

models:
  IndexTeam/IndexTTS-2.5:
    supported_tasks: [voice_clone]
    backend: openai-audio-speech
    endpoint: /v1/audio/speech
    trust_remote_code: true
    task_extra_body:
      voice_clone:
        extra_params:
          lang: en
          text_normalization: true

  # -CustomVoice checkpoints lack speaker_encoder weights, so voice_clone is
  # NOT supported (an attempt raises ValueError from _extract_speaker_embedding
  # at model runtime).  Use -Base for voice_clone.
  Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice:
    supported_tasks: [default_voice, voice_design]
    backend: openai-audio-speech
    endpoint: /v1/audio/speech
    trust_remote_code: true
    task_extra_body:
      default_voice:
        voice: Vivian
        language: English
        task_type: CustomVoice
      voice_design:
        task_type: VoiceDesign
        language: English

  Qwen/Qwen3-TTS-12Hz-1.7B-Base:
    supported_tasks: [voice_clone]
    backend: openai-audio-speech
    endpoint: /v1/audio/speech
    trust_remote_code: true
    task_extra_body:
      voice_clone:
        task_type: Base

  bosonai/higgs-audio-v3-tts-4b:
    supported_tasks: [default_voice, voice_clone]
    backend: openai-audio-speech
    endpoint: /v1/audio/speech
    trust_remote_code: true
    task_extra_body:
      default_voice: {}
      voice_clone: {}

  openbmb/VoxCPM2:
    supported_tasks: [voice_clone, default_voice]
    trust_remote_code: true
    backend: openai-audio-speech
    endpoint: /v1/audio/speech
    trust_remote_code: true
    task_extra_body:
      voice_clone: {}
      default_voice:
        voice: default

  # MOSS-TTS family. After the serving_speech.py variant-detection refactor
  # all five variants now go through OpenAI /v1/audio/speech with the
  # variant-specific request fields (ambient_sound, ref_audio_2, instructions).
  OpenMOSS-Team/MOSS-TTS:
    supported_tasks: [voice_clone]
    backend: openai-audio-speech
    endpoint: /v1/audio/speech
    task_extra_body:
      voice_clone: {}

  # MOSS-TTS-v1.5: continued-training upgrade of MOSS-TTS 1.0 — same
  # MossTTSDelay-8B architecture and API, so identical wiring. Adds 31-language
  # synthesis via the `language` field and inline `[pause Xs]` markers.
  OpenMOSS-Team/MOSS-TTS-v1.5:
    supported_tasks: [voice_clone]
    backend: openai-audio-speech
    endpoint: /v1/audio/speech
    task_extra_body:
      voice_clone: {}

  OpenMOSS-Team/MOSS-TTS-Realtime:
    supported_tasks: [voice_clone]
    backend: openai-audio-speech
    endpoint: /v1/audio/speech
    task_extra_body:
      voice_clone: {}

  # MOSS-TTS-Local-Transformer-v1.5: Qwen3 backbone + 1-layer GPT2-style local
  # depth transformer (n_vq=12), paired with the 48 kHz stereo
  # MOSS-Audio-Tokenizer-v2 codec. Same voice-clone request shape as the
  # other MOSS-TTS variants.
  OpenMOSS-Team/MOSS-TTS-Local-Transformer-v1.5:
    supported_tasks: [voice_clone]
    backend: openai-audio-speech
    endpoint: /v1/audio/speech
    task_extra_body:
      voice_clone: {}

  # MOSS-TTSD synthesizes [S1]/[S2] dialogue. The seed-tts voice_clone dataset
  # has single-speaker rows; bench piggybacks on it (both speakers share the
  # same ref_audio). For real dialogue evaluation use the ttsd dataset.
  OpenMOSS-Team/MOSS-TTSD-v1.0:
    supported_tasks: [voice_clone, dialogue]
    backend: openai-audio-speech
    endpoint: /v1/audio/speech
    task_extra_body:
      voice_clone: {}
      dialogue: {}

  # MOSS-SoundEffect ignores text/ref_audio; expects ambient_sound +
  # duration_seconds. The sound_effect dataset module emits those.
  OpenMOSS-Team/MOSS-SoundEffect:
    supported_tasks: [sound_effect]
    backend: openai-audio-speech
    endpoint: /v1/audio/speech
    task_extra_body:
      sound_effect:
        duration_seconds: 10

  # MOSS-VoiceGenerator: instructions-driven (no ref_audio). seed-tts-design
  # dataset already emits {instructions, target_text}, so voice_design works
  # out of the box once the serving validator accepts instructions.
  OpenMOSS-Team/MOSS-VoiceGenerator:
    supported_tasks: [voice_design]
    backend: openai-audio-speech
    endpoint: /v1/audio/speech
    task_extra_body:
      voice_design: {}