Mistral's Voxtral TTS copies a voice from three seconds

Mistral says Voxtral TTS, released March 23, 2026, can adapt to a new speaker from a reference clip as short as three seconds. It picks up accent, intonation, and even the speaker's hesitations. It does that at 4B parameters, built on Ministral 3B, with a reported model latency of 70ms. Mistral bills it as open weights, and its API starts at $0.016 per 1,000 characters. For anyone building a call center, that is an attractive package. A company can run its own brand voice in nine languages without renting one from a speech provider. The cross-lingual behavior (a French voice prompt with English text yields French-accented English) makes a cascaded translation pipeline cheap to assemble. Mistral's quality claim rests on its own side-by-side test against ElevenLabs Flash v2.5: three annotators, two voices per language, with Voxtral preferred 68.4% of the time on custom voices.

The same three-second trick is exactly what phone fraud needs. Mistral's own translation demo has the agent asking a caller to confirm their full name and date of birth. An attacker with a short clip of an executive, pulled from a conference talk or a voicemail greeting, can now call finance in that voice and ask for a transfer. A closed API can at least add consent checks and watermarks on the server side. Open weights cannot enforce either once downloaded. I still prefer open models, but that preference moves the defense to whoever receives the call. A familiar voice can no longer count as proof of identity: payments and password resets need a callback to a known number or a second channel, whichever speech model the attacker happened to use.