Voice: meeting transcripts no longer drop pieces when the provider is busy
- VoiceA meeting sends a piece of audio every 20 seconds for your microphone and another for the computer audio. Transcription used a single model, which only accepts a few requests per minute, so many pieces failed and the transcript stayed empty. Now, when that model is full, the piece goes to a second transcription model (and to Groq when a key is set), and the full model is skipped for a minute. If every model is full, the recorder waits and sends the piece again instead of losing it. Voice notes in Inbox and Lines and Calls recordings use the same fallback.