feat: enable Whisper STT locally (medium.en)

[jaewilson07]

Summary

• Set whisper.enabled: true in config.yaml — Whisper was installed but disabled
• medium.en model, auto device (uses GPU if available, falls back to CPU)
• config.yaml is volume-mounted so no image rebuild needed; just docker compose restart api

Test plan

• curl http://localhost:7625/v1/config → stt_enabled: true
• Click "start dictating" on the website — voice button enabled, recording indicator appears
• Speak — transcript appears in editor, diagram auto-generates

🤖 Generated with Claude Code

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.

[Copilot]

Pull request overview

Enables local Whisper STT configuration via config.yaml (including medium.en + auto device selection) and improves container resiliency with restart policies.

Changes:

• Move non-secret runtime settings (models, Whisper settings, paths, model filter) from environment variables to backend/config.yaml.
• Add Whisper device selection and handle missing STT dependency imports during app startup.
• Add restart: unless-stopped to ollama and api services in Docker Compose.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 4 comments.

File	Description
voice-to-mermaid/backend/main.py	Loads settings from config.yaml, adds Whisper device config, and attempts graceful handling when STT deps aren’t installed.
voice-to-mermaid/backend/docker-compose.yml	Adds restart policies for ollama and api.
voice-to-mermaid/backend/config.yaml	Defines defaults for Ollama/OpenAI/Whisper and paths; enables Whisper by default.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

If whisper.enabled is true but whisperlivekit isn’t installed, this branch yields/returns but leaves WHISPER_ENABLED as True. Other code paths (e.g. the WebSocket handler’s conditional Whisper imports) can still crash at runtime, and /v1/config will still report stt_enabled: true. Consider introducing a runtime flag like _stt_available (set false on ImportError) or flipping WHISPER_ENABLED to false here and guarding all Whisper imports/usage on that flag.

[Copilot]

These reads assume nested config sections are mappings (e.g. _cfg.get("ollama", {}).get(...)). Since config.yaml is intended to be user-edited, a malformed section like ollama: "..." will raise at import time (AttributeError: 'str' object has no attribute 'get') and prevent startup. Consider validating/coercing each subsection (ollama/openai/whisper/paths) to a dict (or using a small schema/dataclass) before reading nested keys.

Suggested change

	# Secrets — env only
	OLLAMA_URL = os.environ.get("OLLAMA_URL", "http://localhost:11434").rstrip("/")
	OPENAI_URL = os.environ.get("OPENAI_BASE_URL", "").rstrip("/")
	OPENAI_KEY = os.environ.get("OPENAI_API_KEY", "")
	# Settings — config.yaml only; edit config.yaml to change these
	OLLAMA_MODEL = _cfg.get("ollama", {}).get("default_model", "qwen3:8b")
	OPENAI_MODEL = _cfg.get("openai", {}).get("model", "gpt-4o-mini")
	WHISPER_ENABLED = _cfg.get("whisper", {}).get("enabled", False)
	WHISPER_MODEL = _cfg.get("whisper", {}).get("model", "medium.en")
	WHISPER_LANG = _cfg.get("whisper", {}).get("language", "en")
	WHISPER_DEVICE = _cfg.get("whisper", {}).get("device", "auto")
	PROMPT_PATH = Path(_cfg.get("paths", {}).get("prompt", "prompts/mermaid.txt"))
	LOG_DIR = Path(_cfg.get("paths", {}).get("log_dir", "data/logs"))
	_MODEL_FILTER: list[str] = [str(p) for p in (_cfg.get("ollama", {}).get("model_filter") or [])]
	def _config_section(name: str) -> dict:
	section = _cfg.get(name, {})
	if isinstance(section, dict):
	return section
	log.warning("config.%s is not a mapping — using defaults for that section", name)
	return {}
	_ollama_cfg = _config_section("ollama")
	_openai_cfg = _config_section("openai")
	_whisper_cfg = _config_section("whisper")
	_paths_cfg = _config_section("paths")
	# Secrets — env only
	OLLAMA_URL = os.environ.get("OLLAMA_URL", "http://localhost:11434").rstrip("/")
	OPENAI_URL = os.environ.get("OPENAI_BASE_URL", "").rstrip("/")
	OPENAI_KEY = os.environ.get("OPENAI_API_KEY", "")
	# Settings — config.yaml only; edit config.yaml to change these
	OLLAMA_MODEL = _ollama_cfg.get("default_model", "qwen3:8b")
	OPENAI_MODEL = _openai_cfg.get("model", "gpt-4o-mini")
	WHISPER_ENABLED = _whisper_cfg.get("enabled", False)
	WHISPER_MODEL = _whisper_cfg.get("model", "medium.en")
	WHISPER_LANG = _whisper_cfg.get("language", "en")
	WHISPER_DEVICE = _whisper_cfg.get("device", "auto")
	PROMPT_PATH = Path(_paths_cfg.get("prompt", "prompts/mermaid.txt"))
	LOG_DIR = Path(_paths_cfg.get("log_dir", "data/logs"))
	_MODEL_FILTER: list[str] = [str(p) for p in (_ollama_cfg.get("model_filter") or [])]

[Copilot]

whisper.enabled is set to true by default, but the Docker image (Dockerfile installs only requirements.txt) doesn’t include whisperlivekit/faster-whisper. In the default docker compose up flow this will result in STT being effectively unavailable (and currently can still cause runtime import failures unless additionally guarded). Either ship STT deps in the image (e.g., install requirements.stt.txt behind a build arg/profile) or default whisper.enabled to false and document the steps needed to enable STT locally.

Suggested change

	enabled: true
	enabled: false # Enable only when whisperlivekit/faster-whisper dependencies are installed in the environment/image.

[Copilot]

By switching these settings to config.yaml only, existing setup docs/samples that instruct users to set .env values like OLLAMA_MODEL, OPENAI_MODEL, WHISPER_MODEL/DEVICE, MERMAID_PROMPT_PATH, and model filtering will now be incorrect (those env vars are ignored). If this is intentional, please update README.md and backend/.env.example accordingly (or reintroduce env overrides) to avoid confusing deployments.