Transcription: move to an OpenAI-compatible cheaper provider (Groq whisper-large-v3-turbo) with a domain vocabulary prompt #297

New Issue

2026-10-27T20:03:49-04:00

logan commented

2026-10-27 20:03:49 -04:00

From the 2026-10-27 review (#296). whisper-1 costs about $1.50 per 1,000 calls; Groq turbo costs about $0.17 and speaks the same API. Off-the-shelf Whisper scores about 51% WER on police radio (arXiv 2409.10858), so the prompt (224 tokens: local streets, unit IDs, 10-codes) matters more than the model.

Make the provider, base URL and model config-driven, keep OpenAI as the fallback, and feed the talkgroup/area vocabulary into the prompt. Later: fine-tune on corrected transcripts (#163). Section issues: #162, #83, #36, #48.

From the 2026-10-27 review (#296). whisper-1 costs about $1.50 per 1,000 calls; Groq turbo costs about $0.17 and speaks the same API. Off-the-shelf Whisper scores about 51% WER on police radio (arXiv 2409.10858), so the prompt (224 tokens: local streets, unit IDs, 10-codes) matters more than the model. Make the provider, base URL and model config-driven, keep OpenAI as the fallback, and feed the talkgroup/area vocabulary into the prompt. Later: fine-tune on corrected transcripts (#163). Section issues: #162, #83, #36, #48.

logan added the

type

feature

area/ai-pipeline area/transcription

priority

p1

labels 2026-10-27 20:03:49 -04:00

logan commented

2026-11-08 21:51:07 -05:00