Telnyx AI Assistants can now run more than one transcription model on the same call. Two new modes, Fallback and Language Booster, keep calls going when a transcription model fails and pick the most accurate transcript for every turn.
transcription.fallback_models. best_turn (default) picks the better transcript each turn, best_engine locks onto the stronger model for the call, and merge_words merges Arabic and English spoken in the same sentence. Set through transcription.challenger.
An assistant runs one mode at a time, and each backup model or booster carries its own language and provider settings.
A voice agent is only as good as the transcript it reasons on. Until now you chose one transcription model for every call: if it failed, the call degraded with it, and if callers switched languages mid-sentence, accuracy dropped. With both modes on one control plane, you compose the right pair of models for each deployment instead of adding another vendor or another integration. In our internal tests, deepgram/flux with deepgram/nova-3 as the booster cut English word errors from 5.7% to 4.6%, and the Basira and Cohere merge_words pair cut errors from 27.7% to 22.4%.
"transcription": {
"model": "deepgram/flux",
"language": "en",
"fallback_models": [
{ "model": "deepgram/nova-3" },
{ "model": "soniox/stt-rt-v5" }
]
}
Learn more in the transcription settings docs or on the Voice AI Agents page.