Add speaker count controls, mic mixing, and configurable cloud transcription

[Copilot]

The recorder was over-splitting a two-person meeting into extra speaker labels, loopback capture did not reliably include the local speaker’s microphone, and local Whisper startup/resource usage made low-spec machines difficult to support.

• Speaker diarization controls

  • Added exact speaker count support via --speaker-count / --speakers.
  • Added --max-speakers to cap diarization when the exact count is unknown.
  • Passes speaker constraints through to pyannote and applies the same cap to the energy-based fallback.

• Local microphone capture

  • Added optional microphone mixing with WASAPI loopback via --include-mic.
  • Added --mic-device and --mic-gain for device selection and level tuning.
  • Saves the mixed audio when microphone capture is enabled, so playback and transcript inputs match.

• Configurable cloud transcription provider

  • Added API-backed transcription without loading local Whisper models.
  • Supports --provider openai, --provider vercel, and --provider compatible.
  • Supports configurable cloud models via --transcription-model / TRANSCRIPTION_MODEL.
  • Supports Vercel AI Gateway via AI_GATEWAY_API_KEY.
  • Supports generic OpenAI-compatible transcription APIs via --transcription-base-url / TRANSCRIPTION_BASE_URL.
  • Keeps local faster-whisper as the default provider.
  • Keeps --openai-model, OPENAI_API_KEY, and OPENAI_TRANSCRIBE_MODEL working for backwards compatibility.

• Docs and coverage

  • Documented recommended usage for two-speaker meetings, microphone mixing, OpenAI cloud transcription, Vercel AI Gateway, and compatible cloud transcription endpoints.
  • Added tests for speaker caps, cloud provider selection, endpoint construction, model fallback behavior, config env fallback behavior, and transcription request handling.

Example usage:

# Two-person meeting: prevent extra Speaker 3/4 labels
python recorder.py start --speaker-count 2

# Capture both remote audio and your microphone
python recorder.py start --include-mic --mic-gain 1.5

# Use OpenAI cloud transcription on low-resource systems
set OPENAI_API_KEY=sk_your_key_here
python recorder.py start --provider openai --transcription-model whisper-1

# Use Vercel AI Gateway / compatible model routing
set AI_GATEWAY_API_KEY=your_gateway_key_here
python recorder.py start --provider vercel --transcription-model openai/whisper-1

# Use another OpenAI-compatible transcription endpoint
set TRANSCRIPTION_API_KEY=your_key_here
python recorder.py start --provider compatible --transcription-base-url https://example.com/v1 --transcription-model provider/model

[buzz39]

@copilot can you add vercel SDK so we can use any model instead of only hardocing to openai

[Copilot]

@copilot can you add vercel SDK so we can use any model instead of only hardocing to openai

Addressed in 8825eb3. Cloud transcription is no longer hardcoded to OpenAI: it now supports --provider vercel for Vercel AI Gateway plus --provider compatible with --transcription-base-url and --transcription-model for other OpenAI-compatible transcription APIs.