From 6ad852692010d146298fe8dc2ebef26f2ff31c7f Mon Sep 17 00:00:00 2001 From: Felix Martin Date: Sat, 22 Aug 2026 09:22:13 -0400 Subject: [PATCH] Add Voxtral Small 24B as a transcription option Whisper large-v3 mis-segments unfamiliar proper nouns: "Whisper Large" came back as "Wisp Alarge", and it leaves doubled words and sentence breaks in the wrong places. Voxtral decodes with a 24B language model, so ambiguous acoustics get resolved against what a sentence can mean. Checked both against saved captures: Voxtral fixed the proper noun and the punctuation, at 1.7 s round-trip for 15 s of audio. Co-Authored-By: Claude Opus 5 (1M context) Claude-Session: https://claude.ai/code/session_01QrC2xo3RUExhVPzWGnGfRK --- blurt.py | 5 +++++ 1 file changed, 5 insertions(+) diff --git a/blurt.py b/blurt.py index 7830e83..3f908af 100644 --- a/blurt.py +++ b/blurt.py @@ -63,6 +63,11 @@ TRANSCRIPTION_MODELS = [ "OpenRouter: gpt-4o-transcribe", "openrouter", ), + ( + "mistralai/voxtral-small-24b-2507-stt", + "OpenRouter: voxtral-small-24b (LLM-based)", + "openrouter", + ), ] PROVIDER_KEY_HINTS = { "openai": ("set OPENAI_API_KEY or configure keyring/Credential Manager entry blurt/openai-api-key"),