AI voice models: Mistral Voxtral vs Microsoft MAI-Voice

AI voice models, episode 6. Previous: Kokoro vs Chatterbox vs Fish Audio Drama 3. Next: the wrap-up comparison table.

Two big platforms that are not voice specialists now sell their own speech models. Mistral shipped Voxtral Mini TTS 3 with Voxtral 3 on 20 October, alongside new transcription models and a voice agent SDK. Microsoft AI’s in-house lab sells MAI-Voice-2.1 and a faster MAI-Voice-2.1-Flash through Microsoft Foundry and Azure Speech. Both pitch voice agents, and both undercut the specialists on price.

The two against the checklist

1. Voice cloning. Mistral offers voice cloning in its Studio playgrounds and a higher tier, Pro Voice, that it says reproduces a speaker’s character and emotional range across languages from 30 minutes of audio. The older open Voxtral 4B TTS card benchmarks with a 10-second reference clip and says only that you are responsible for lawful use. Microsoft’s MAI-Voice-2 does zero-shot voice prompting from 5 to 60 seconds of reference audio, but with a hard gate: “only authorized, licensed voices can be synthesized in production”, and you apply for access. Azure Speech also has its older professional voice fine-tuning and personal voice options.

2. Languages. Mistral says Voxtral Mini TTS 3 covers 35+ languages and 50+ dialects, with preset voices in 18 languages. MAI-Voice-2.1 lists 23 languages, up from 15 at MAI-Voice-2 and English only for MAI-Voice-1. Microsoft also notes code-switching for pairs such as Hindi-English and Spanish-English.

3. Latency and streaming. Microsoft lists model inference latency of about 550 ms for MAI-Voice-2.1 and about 45 ms for Flash. Mistral offers batch and realtime TTS and says its agent SDK completes each conversational turn in under one second; the open 4B model’s card shows 70 ms latency at one request on an H200. These are vendor figures measured in different ways.

4. Emotion and style. Voxtral picks emotion, pace and emphasis from the conversation so far and follows custom prompt instructions. MAI-Voice uses emotion tags (sad, whispered, excited and so on) for granular control.

5. Output and quality. The open Voxtral 4B TTS writes 24 kHz audio in WAV, PCM, FLAC, MP3, AAC or Opus. Neither Mistral’s Voxtral 3 post nor Microsoft’s MAI-Voice pages we read state an output sample rate for the current models.

6. Price. Voxtral Mini TTS 3 is $13 per million characters. MAI-Voice-2.1 is $22, and Flash $15. For scale, Artificial Analysis lists ElevenLabs’ top model, Eleven v4, at $80.

7. Speech-to-speech and realtime. Neither TTS model is speech-to-speech on its own. Mistral’s answer is the LiveKit-based agent SDK, a cascade of realtime transcription, a general model and Voxtral TTS, with SIP and WebRTC support. Microsoft’s is Azure Voice Live, which builds speech-to-speech agents on MAI-Voice-2-Flash, and Dynamics 365 Contact Center, which Microsoft says now runs on Flash. Our voice agent pipeline explainer covers the trade-offs of that cascaded design.

8. Long-form. Microsoft claims stable speaker identity across audiobooks, podcasts and lectures, and says 2.1 holds up across long generations. Mistral’s post does not make a specific long-form claim.

9. Against the previous version. On the Artificial Analysis board, Mistral’s March Voxtral TTS sits 50th with an Elo of 1092; Voxtral 3 TTS is 17th at 1215, a large jump. Microsoft says MAI-Voice-2 was preferred over its predecessors 72% of the time in its own tests, and that MAI-Voice-2-Flash is twice as fast as MAI-Voice-2 and 32% cheaper.

10. What users think. Independent data is thin for Microsoft: MAI-Voice-1 is 81st on Artificial Analysis at 1028, and MAI-Voice-2.x does not appear on the board we read. Microsoft points to production use instead, saying MAI-Voice-2-Flash cut GPU costs by up to 89% in Dynamics 365 Contact Center, used by customers such as T-Mobile and EasyJet. Mistral’s quality charts are its own and were scored by a Gemini model as judge.

11. Licence and deployment. Voxtral 3 runs on Mistral’s EU, US and global API endpoints, and Mistral says it can be self-deployed, without detail in the post. The only open Voxtral TTS weights we found are the 4B model, under CC BY-NC 4.0, so not for commercial use. MAI-Voice is closed and available through Azure; we found no weights.

12. Verdict. Voxtral is the cheaper pick, with more languages, an agent SDK and EU hosting, and it now ranks well on an independent board. MAI-Voice suits teams already on Azure that want consent-gated cloning and Flash’s low latency, but its newer models still lack independent rankings. Run your own scripts through both before you choose.

Also worth knowing

• Mistral’s transcription side of Voxtral 3 starts at $0.0035 per minute for the 4B Mini model, which matters if you price a whole agent, not only the voice.
• Azure Speech still sells standard and HD neural voices alongside MAI-Voice, so check which model a given Azure voice actually uses.
• For xAI’s voice stack, see episode 3 on Grok Voice.

Why this matters

The voice layer is getting cheap: both platforms now sit well under the $80 per million characters Artificial Analysis lists for Eleven v4. The real differences are where your audio is processed, whether cloning is gated, and whether you need a non-commercial open model for testing. If you build a voice agent, price the full turn, transcription plus model plus voice, and keep a consent record for any cloned voice even where the platform enforces one.

Summary

	Voxtral Mini TTS 3	MAI-Voice-2.1 / Flash
Price per 1M chars	$13	$22 / $15
Languages	35+ (presets in 18)	23
Latency (vendor)	Agent turn under 1 s	~550 ms / ~45 ms inference
Cloning	Studio cloning; Pro Voice from 30 min	5-60 s prompt, authorized voices only
Realtime agents	LiveKit agent SDK, SIP, WebRTC	Azure Voice Live, Dynamics 365
AA Elo (10 Nov)	1215, 17th	Not listed (MAI-Voice-1: 1028, 81st)
Open weights	Older 4B model, CC BY-NC	None found

We did not generate our own samples for this review. Both vendors’ pages link official demos.