[FEATURE] FluidMeet: offer Cohere Transcribe for meeting transcripts #1278

Problem

FluidMeet only accepts Parakeet models for meetings (v2, v3 and Fluid Speech Pro). FluidVoice already ships Cohere Transcribe for dictation, and of the FluidVoice models scored on the Open ASR Leaderboard's set of real meeting recordings (AMI, 2 November version), it gets the fewest words wrong:

Cohere gets about 25% fewer meeting words wrong than v3, which FluidMeet uses since 1.6.13-beta.2 when it's the dictation model. The leaderboard tests full-precision models, and FluidVoice ships Cohere at 6-bit, so the gap on a Mac may be smaller.

Proposed solution

Add a speech model setting to FluidMeet that includes Cohere Transcribe, with Parakeet staying the default since it's much faster.

Cohere would first need the fix in #736. A commenter there reports each audio piece is capped at 96 tokens, so fast speech loses the end of a piece, and an hour-long meeting is over a hundred pieces.

Alternatives considered

Switching dictation to Cohere doesn't help, because FluidMeet falls back to a Parakeet model when the dictation model isn't one it supports.