---
# Audio Model Configuration Matrix
# Maps audio models to test scenarios for vLLM CPU performance evaluation

audio_models:
  # Whisper Models (ASR - Automatic Speech Recognition)
  - name: "whisper-tiny"
    model_id: "openai/whisper-tiny"
    task_type: "transcription"
    description: "Fastest Whisper model, 39M parameters, good for quick transcription"
    supported_endpoints:
      - "/v1/audio/transcriptions"
      - "/v1/audio/translations"
    test_scenarios:
      - "transcription-throughput"
      - "transcription-latency"
      - "audio-duration-scaling"
      - "constant-rate-stress"
      - "format-comparison"
      - "transcription-quality"
      - "quick-test"
    recommended_settings:
      dtype: "float16"
      max_model_len: 448

  - name: "whisper-small"
    model_id: "openai/whisper-small"
    task_type: "transcription"
    description: "Small Whisper model, 244M parameters, balanced speed/quality"
    supported_endpoints:
      - "/v1/audio/transcriptions"
      - "/v1/audio/translations"
    test_scenarios:
      - "transcription-throughput"
      - "transcription-latency"
      - "audio-duration-scaling"
      - "constant-rate-stress"
      - "format-comparison"
      - "transcription-quality"
      - "quick-test"
    recommended_settings:
      dtype: "float16"
      max_model_len: 448

  - name: "whisper-medium"
    model_id: "openai/whisper-medium"
    task_type: "transcription"
    description: "Medium Whisper model, 769M parameters, high quality transcription"
    supported_endpoints:
      - "/v1/audio/transcriptions"
      - "/v1/audio/translations"
    test_scenarios:
      - "transcription-throughput"
      - "audio-duration-scaling"
      - "quick-test"
    recommended_settings:
      dtype: "bfloat16"
      max_model_len: 448

  # Audio Chat Models (future - no runnable scenarios yet)
  - name: "ultravox-v0_5-llama-3_2-1b"
    model_id: "fixie-ai/ultravox-v0_5-llama-3_2-1b"
    task_type: "audio_chat"
    description: "Multimodal audio chat model, processes audio + text (scenarios not yet implemented)"
    supported_endpoints:
      - "/v1/chat/completions"
    test_scenarios: []
    recommended_settings:
      dtype: "bfloat16"
      max_model_len: 2048

# Common test datasets for audio benchmarking
datasets:
  - name: "librispeech_asr_clean"
    dataset_id: "openslr/librispeech_asr"
    config: "clean"
    split: "test"
    description: "Clean English speech, high quality, good for ASR baseline"
    audio_column: "audio"
    typical_duration: "5-15 seconds"
    language: "en"

  - name: "librispeech_asr_other"
    dataset_id: "openslr/librispeech_asr"
    config: "other"
    split: "test"
    description: "More challenging English speech with background noise"
    audio_column: "audio"
    typical_duration: "5-15 seconds"
    language: "en"

  - name: "common_voice_en"
    dataset_id: "mozilla-foundation/common_voice_11_0"
    config: "en"
    split: "test"
    description: "Diverse speakers and accents, real-world audio"
    audio_column: "audio"
    typical_duration: "3-10 seconds"
    language: "en"

# Audio preprocessing configurations
preprocessing:
  formats:
    - name: "mp3_64k"
      format: "mp3"
      bitrate: "64k"
      sample_rate: 16000
      mono: true
      description: "Compressed audio, bandwidth-efficient"

    - name: "wav_16k"
      format: "wav"
      sample_rate: 16000
      mono: true
      description: "Uncompressed audio, high quality"

    - name: "flac_16k"
      format: "flac"
      sample_rate: 16000
      mono: true
      description: "Lossless compression, balanced quality/size"

  duration_buckets:
    - name: "short"
      max_duration: 5.0
      description: "Short utterances (1-5 seconds)"

    - name: "medium"
      max_duration: 30.0
      description: "Medium clips (5-30 seconds)"

    - name: "long"
      max_duration: 120.0
      description: "Long recordings (30-120 seconds)"

    - name: "unlimited"
      max_duration: null
      description: "Full audio, no truncation"