Mistral's Voxtral 3 bundles transcription, speech and a voice agent SDK, and can run on your own servers. Filipino is in its transcription list

Voice agents keep getting easier to assemble. The hard part is still what they are allowed to do once the caller says yes.

What happened

Mistral launched Voxtral 3, a family of speech models plus a voice agent SDK built on LiveKit. The stack includes Voxtral Small Transcribe 3 for speech to text, Voxtral Mini TTS 3 for speech, and an SDK that runs listening, reasoning, tool calling and speaking as one streaming pipeline. Mistral says each turn completes in under one second and the agent stops talking when the caller cuts in.

Transcription covers 50+ languages, and Filipino appears in the language lists Mistral published for its benchmarks. Speech output supports 35+ languages and 50+ dialects, though the ready made voice presets cover 18 languages and Filipino is not one of them. Mistral lists transcription on the Mini variant at $0.0035 per minute and speech at $13 per million characters. It is available through Mistral's EU, US and global API endpoints or for self deployment, which lets a company control where audio is processed and whether it is kept.

Mistral's own examples include agents that book appointments and check invoices during a live call.

My take

For operators, the model quality is now the easy part. What decides whether a voice agent helps a business is the plumbing around it:

1. Which tools the agent can call, and with which permissions. Booking a slot is fine. Issuing a refund probably needs a human.
2. What gets written to the CRM after the call: a summary, the outcome, the next step, and the recording only if the client agreed to keep it.
3. A clear handoff path when the caller is confused or angry.

Self hosting matters for clients in regulated work like clinics or finance, where audio leaving their environment is a dealbreaker. For Philippine based support teams, test transcription on real calls with Taglish before you commit. A benchmark language list is a starting point, not proof it handles your customers.

I would start with one narrow call type, like appointment confirmations, and expand only when the logs look clean.