Skip to main content

MicrosoftAudio modelActive

MAI-Voice-2

Popularityno data yet

n/ano search or usage data

Summary

MAI-Voice-2 is Microsoft’s earlier Foundry speech synthesis model with 15 supported languages. A 5 to 60 second reference recording can guide voice cloning, with Microsoft describing consent guardrails.

Microsoft released MAI-Voice-2 on June 2, 2026. Its Foundry model identifier for audio output is mai-voice-2.

GraduallyBench Audio

In the speech output text comparison, the re-transcription of MAI-Voice-2 differs from the source by 3.4 % (11 errors over 321 words, German and English). The median of the 20 models with a German and English value is 3.5 %.

Word error rate, German and English

3.4 %

11 errors over 321 reference words

Character error rate

2.0 %

33 errors over 1,685 characters

Scope

16 / 16

valid of planned outputs

Access

OpenRouter

Text comparison, German and English, compared

Differing words after re-transcription in percent (German and English), lower is better

20 of 20 models

Smaller test scope: Fish Audio S2 Pro, Fish Audio S2.1 Pro, MAI-Voice-2, MAI-Voice-2-Flash, Sonic 3.6, Voxtral Mini TTS, GPT-4o Mini TTS.

As of 11/10/2026. Source: GraduallyBench Audio

MAI-Voice-2

Benchmark run
  • The text comparison uses a second speech recognizer as a proxy.
  • Naturalness and voice quality were not evaluated.
  • The voice actually used is not confirmed.
  • The route does not confirm the exact model version.
Show test details

Speech output with text comparison

OpenRouter · German, English

16 valid outputs from 16 attempted and 16 planned test slots.

Speech output text comparison, WER
2.48 %(4 / 161 reference words)German
Speech output text comparison, CER
1.01 %(9 / 891 reference characters)German
Speech output text comparison, WER
4.38 %(7 / 160 reference words)English
Speech output text comparison, CER
3.02 %(24 / 794 reference characters)English
Speech output text comparison, WER
3.43 %(11 / 321 reference words)German, English
Speech output text comparison, CER
1.96 %(33 / 1,685 reference characters)German, English

WER compares words and CER compares characters with the reference. The speech output was re-transcribed with GPT Transcribe. Its exact backend revision is unconfirmed. Read speech and digital silence do not cover spontaneous speech or real background noise. Error rates can exceed 100 %. Where used, the test references were not independently reviewed by humans. Test coverage and methods differ by model. Each metric stands on its own; there is no combined overall score across tests. Data compiled on 11/10/2026. How we test · Download all Decision and Audio results as JSON

All benchmark results

Every published result with its source, test setup and comparison models.

Show all 6 published benchmark results

Measurements without matching peer values

These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.

win rate versus Voice 1 (Show measurement, test conditions, and source)
Source value
72.1
Metric
win rate
Unit
%
Benchmark version
source version not stated
Category
audio
Direction
Higher is better
Source type
vendor-reported
Evaluator
Microsoft MAI-Voice-2 announcement
Status
active
Retrieved at
2026-10-04
Methodology
Vendor-reported result. 2,500 listening tests; 2,222 responses; 11 languages; preference test, not MOS
Context
Source-specific observation; it does not establish a compatible cross-model cohort.
Source Microsoft MAI-Voice-2 announcement
Microsoft listener preference versus human recordings (Show measurement, test conditions, and source)
Source value
45.5
Metric
share of responses preferring MAI-Voice-2 speech
Unit
%
Benchmark version
Microsoft announcement listening-test chart, 11 languages; source version not stated
Category
audio
Direction
Higher is better
Sample
2222
Source type
vendor-reported
Evaluator
Microsoft AI
Status
active
Retrieved at
2026-10-16
Methodology
Vendor-reported result. Across 11 languages and 2,222 responses, Microsoft reports 45.5% preferring generated MAI-Voice-2 speech, 44% preferring real human recordings, and 10.5% tie. This is a preference test, not MOS.
Context
Vendor-reported preference result. It must not be interpreted as MOS, an absolute naturalness score, or a direct model-versus-model result. The existing 72.1% comparison with MAI-Voice-1 is a separate cohort and observation.
Source Microsoft AI
Design Arena Audio Realism public leaderboard Elo, 2026-10-17 (Show measurement, test conditions, and source)
Source value
1,209
Metric
Elo-style rating from Bradley-Terry pairwise preference votes
Unit
elo
Benchmark version
Design Arena Audio Realism public leaderboard snapshot 2026-10-17
Category
audio realism
Direction
Higher is better
Source type
official-leaderboard
Evaluator
Design Arena Audio Realism leaderboard
Status
superseded
Retrieved at
2026-10-17
Methodology
Official benchmark leaderboard. The public Audio Realism benchmark compares two recordings of the same prompt blind, with randomized order; vetted native speakers vote for the voice that sounds more human. Bradley-Terry aggregation produces the Elo-style rating. It uses 500+ fixed, versioned private prompts in phone-agent, explainer, and conversation categories. One male and one female American-English voice are selected per model, and matchups use the same prompt and gender. The 2026-10-17 API row reports 385 wins, 243 losses, and 628 battles for exact model ID mai-voice-2.
Context
The current API reports 628 pairwise battles but no confidence interval or effective independent sample size; the vote count is not stored as sampleCount. The article's 1224 Elo is an older published snapshot, not an independent replication.
Source Design Arena Audio Realism leaderboard
Cantari Open AI Voice Benchmark, routed latency (Show measurement, test conditions, and source)
Source value
2,426
Metric
wall-clock time from request to receipt of full audio bytes
Unit
ms
Category
audio-latency
Direction
Lower is better
Source type
independent-evaluator
Evaluator
Cantari
Status
active
Retrieved at
2026-10-22
Methodology
Independent evaluation. Cantari sends the same English passage through the routed speech endpoint used by its studio, records three runs, and publishes the median time until the full audio bytes arrive.
Context
This is latency on Cantari's routed path, not a service-wide SLA or a quality score. Cantari explicitly says the displayed MAI Voice 2 Elo belongs to MAI-Voice-1; MAI Voice 2 is not arena-rated there. The serving revision is not specified.
Source Cantari
DesignArena Audio Realism Elo (Show measurement, test conditions, and source)
Source value
1,208
Metric
Bradley-Terry preference Elo for perceived human-like voice realism
Unit
elo
Benchmark version
DesignArena Audio Realism public board snapshot 2026-10-28
Direction
Higher is better
Source type
independent-evaluator
Evaluator
DesignArena
Status
active
Retrieved at
2026-10-28
Methodology
Independent evaluation. Current public Audio Realism row mai-voice-2; one male and one female American-English voice selected per model; blind same-prompt, same-gender comparisons using a fixed held-out set of 500+ versioned prompts across phone-agent, explainer and conversation categories; 640 battles.
Source DesignArena
DesignArena TTS Arena Elo (Show measurement, test conditions, and source)
Source value
1,160
Metric
blind pairwise preference Elo
Unit
elo
Benchmark version
DesignArena public TTS board snapshot 2026-10-28
Direction
Higher is better
Source type
independent-evaluator
Evaluator
DesignArena
Status
active
Retrieved at
2026-10-28
Methodology
Independent evaluation. Current DesignArena public TTS row mai-voice-2; same-prompt pairwise human preference; 2582 battles. Board update 2026-10-28 16:06 UTC.
Context
Battles are pairwise-comparison appearances, not independent prompt count; no row-level confidence interval published.
Source DesignArena

All specifications

Architecture

Model class
Speech synthesis
Source: Model class

Input and output

Input
Text and reference audio
Source: Input
Output
Audio
Source: Output
Languages
15 languages
Source: Languages

Watermarking and provenance

AI watermarking refers to invisible signals embedded in the output. Visible marks and provenance metadata are listed separately. Unknown means there is insufficient evidence for that particular field.

AI watermarking
Unknown. Microsoft lists this model for content provenance in Foundry. Its table does not establish whether an invisible watermark is used.
Source: AI watermarking
Watermarking technique
Not known
Source: Watermarking technique
Labelling reviewed
Nov 7, 2026
Source: Labelling reviewed

Licence and access

Access
Microsoft Foundry API
Source: Access
Shutdown
None announced
no source

API

API model ID
mai-voice-2
Source: API model ID

Further details

Voice cloning
A 5 to 60 second reference recording with consent guardrails
Source: Voice cloning

Sources and data date

Show all 9 sources

Benchmark sources at a glance

Which independent evaluators tested this model. This is coverage, not a quality ranking.

No active result from the reviewed external evaluators can currently be attributed to this model. Other source evidence remains available below.

Without vendor claims and historical or disputed results; those stay labelled in the single values.