Audio pipeline spec template

Reference template for the audio pipeline. Every encode strategy and every generate option is documented inline. Copy this into ~/.config/mm/pipelines/audio/{mode}.yaml (replacing {mode} with fast or accurate) and edit only the fields you want to override — omitted keys fall back to the built-in defaults.

Source: python/mm/pipelines/audio/spec.yaml.template

# Audio pipeline spec — all available strategies and their options.
#
# Copy this file to ~/.config/mm/pipelines/audio/{mode}.yaml
# and customise the values you need.  Omitted strategy_opts fall back
# to their defaults.  The generate block is optional — leave it out
# for encode-only (transcript-only) pipelines.

kind: audio
mode: fast        # fast | accurate

# ── encode ──────────────────────────────────────────────────────────

encode:

  # ── transcribe (default) ──────────────────────────────────
  # Transcribes audio via the modular backend system.  By default
  # uses the VLM Run gateway (OpenAI-compatible).  Override base_url
  # to point at localhost or OpenAI directly.
  strategy: transcribe
  # model: nvidia/parakeet-tdt-0.6b-v3     # encoder model (default for gateway)
  # backend: openai                        # "openai" | "mlx" | "ctranslate2"
  strategy_opts:
    # language: auto                        # language code or "auto"
    # audio_speed: 2.0                      # playback speed multiplier
    # base_url: null                        # custom OpenAI-compatible endpoint
    # api_key: null                         # API key for that endpoint

  # ── base64 ────────────────────────────────────────────────
  # Sends raw audio as a base64 input_audio part (no transcription).
  # Best for models with native audio understanding.
  # strategy: base64
  # strategy_opts:
  #   format: mp3             # mp3 | wav | flac | ogg | m4a | aac | opus | webm
  #                           # default: inferred from file extension

  # ── gemini ────────────────────────────────────────────────
  # Passes audio as Gemini inline_data Part(s).  Long files are
  # automatically chunked.
  # strategy: gemini
  # strategy_opts:
  #   max_seconds: 120        # max chunk length in seconds
  #   overlap: 10             # overlap between chunks in seconds

  # Custom post-processing transform (optional).
  # pyfunc: null

# ── generate (optional — omit for encode-only) ──────────────────

generate:
  prompt: >-
    Summarize this audio recording ({filename}) based on the transcript.
    Provide a concise summary (~80 words) covering the main topics discussed.

    Transcript:
    {transcript}
  max_tokens: 512
  # model: null              # pin a specific LLM model (overrides profile)
  # temperature: null        # sampling temperature (null = model default)
  # json_mode: false         # request JSON-formatted response
  # think: false             # enable extended thinking
  # reasoning_effort: none   # none | low | medium | high
  # extra_body: {}           # provider-specific pass-through kwargs