MicrosoftAudio modelPreview

MAI-Voice-2.1-Flash

PopularityRank 654 of 875

0out of 100-1 vs. last month

Summary

MAI-Voice-2.1-Flash is the low-latency variant of Microsoft’s current speech synthesis line. It supports the same 23 languages and 26 locales and can generate up to 45 seconds of audio in one generation.

Microsoft cites about 150 milliseconds end to end and a price of $15 per 1M characters. The Foundry API identifier is mai-voice-2-1-flash. Access is a public preview without an SLA, and the weights are not public.

All benchmark results

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

Show all 8 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.

Cekura TTS hard-part accuracy: numbers and shorthand (Show measurement, test conditions, and source)

Source value: 79
Score: 79
Metric: Share of hard-part readings accepted for numbers and shorthand
Unit: %
Benchmark version: Cekura TTS Model Benchmarks; current page snapshot 2026-10-19; run date and serving revision not stated
Category: audio
Direction: Higher is better
Source type: official-leaderboard
Evaluator: Cekura Bench TTS Model Benchmarks
Status: active
Retrieved at: 2026-10-19
Methodology: Official benchmark leaderboard. Evaluator-run TTS benchmark. Cekura says each model read the same 104 agent sentences three times from one US East cloud region, using one stock voice and its own streaming protocol. Non-streaming timing runs from sending the full text to first audible sound; streaming timing runs from the last word sent to first audible sound. Reading accuracy is estimated from Deepgram Nova-3 and OpenAI Whisper transcripts, counting accepted readings without human review. Naturalness uses UTMOS, an open model predicting a listening-test score. The captured rows do not state per-cell N, CI, exact run date, serving/API revision, or exact voice ID unless shown separately.
Context: Cekura discloses that it runs the benchmark itself and sells testing and monitoring tools in the same market. The results are evaluator-published and include voice, runtime, protocol, ASR, and UTMOS effects; they are not a human-only or isolated-model evaluation.

Source Cekura Bench TTS Model Benchmarks

Cekura TTS hard-part accuracy: codes and addresses (Show measurement, test conditions, and source)

Source value: 67
Score: 67
Metric: Share of hard-part readings accepted for codes and addresses
Unit: %
Benchmark version: Cekura TTS Model Benchmarks; current page snapshot 2026-10-19; run date and serving revision not stated
Category: audio
Direction: Higher is better
Source type: official-leaderboard
Evaluator: Cekura Bench TTS Model Benchmarks
Status: active
Retrieved at: 2026-10-19
Methodology: Official benchmark leaderboard. Evaluator-run TTS benchmark. Cekura says each model read the same 104 agent sentences three times from one US East cloud region, using one stock voice and its own streaming protocol. Non-streaming timing runs from sending the full text to first audible sound; streaming timing runs from the last word sent to first audible sound. Reading accuracy is estimated from Deepgram Nova-3 and OpenAI Whisper transcripts, counting accepted readings without human review. Naturalness uses UTMOS, an open model predicting a listening-test score. The captured rows do not state per-cell N, CI, exact run date, serving/API revision, or exact voice ID unless shown separately.
Context: Cekura discloses that it runs the benchmark itself and sells testing and monitoring tools in the same market. The results are evaluator-published and include voice, runtime, protocol, ASR, and UTMOS effects; they are not a human-only or isolated-model evaluation.

Source Cekura Bench TTS Model Benchmarks

Cekura TTS hard-part accuracy: read from context (Show measurement, test conditions, and source)

Source value: 75
Score: 75
Metric: Share of hard-part readings accepted for context-dependent readings
Unit: %
Benchmark version: Cekura TTS Model Benchmarks; current page snapshot 2026-10-19; run date and serving revision not stated
Category: audio
Direction: Higher is better
Source type: official-leaderboard
Evaluator: Cekura Bench TTS Model Benchmarks
Status: active
Retrieved at: 2026-10-19
Methodology: Official benchmark leaderboard. Evaluator-run TTS benchmark. Cekura says each model read the same 104 agent sentences three times from one US East cloud region, using one stock voice and its own streaming protocol. Non-streaming timing runs from sending the full text to first audible sound; streaming timing runs from the last word sent to first audible sound. Reading accuracy is estimated from Deepgram Nova-3 and OpenAI Whisper transcripts, counting accepted readings without human review. Naturalness uses UTMOS, an open model predicting a listening-test score. The captured rows do not state per-cell N, CI, exact run date, serving/API revision, or exact voice ID unless shown separately.
Context: Cekura discloses that it runs the benchmark itself and sells testing and monitoring tools in the same market. The results are evaluator-published and include voice, runtime, protocol, ASR, and UTMOS effects; they are not a human-only or isolated-model evaluation.

Source Cekura Bench TTS Model Benchmarks

Cekura TTS hard-part accuracy: medical terms (Show measurement, test conditions, and source)

Source value: 92
Score: 92
Metric: Share of hard-part readings accepted for medical terms
Unit: %
Benchmark version: Cekura TTS Model Benchmarks; current page snapshot 2026-10-19; run date and serving revision not stated
Category: audio
Direction: Higher is better
Source type: official-leaderboard
Evaluator: Cekura Bench TTS Model Benchmarks
Status: active
Retrieved at: 2026-10-19
Methodology: Official benchmark leaderboard. Evaluator-run TTS benchmark. Cekura says each model read the same 104 agent sentences three times from one US East cloud region, using one stock voice and its own streaming protocol. Non-streaming timing runs from sending the full text to first audible sound; streaming timing runs from the last word sent to first audible sound. Reading accuracy is estimated from Deepgram Nova-3 and OpenAI Whisper transcripts, counting accepted readings without human review. Naturalness uses UTMOS, an open model predicting a listening-test score. The captured rows do not state per-cell N, CI, exact run date, serving/API revision, or exact voice ID unless shown separately.
Context: Cekura discloses that it runs the benchmark itself and sells testing and monitoring tools in the same market. The results are evaluator-published and include voice, runtime, protocol, ASR, and UTMOS effects; they are not a human-only or isolated-model evaluation.

Source Cekura Bench TTS Model Benchmarks

Cekura TTS hard-part accuracy estimate (Show measurement, test conditions, and source)

Source value: 78.3
Score: 78.3
Metric: Share of hard parts read correctly, estimated from speech-to-text transcripts
Unit: %
Benchmark version: Cekura TTS Model Benchmarks; current page snapshot 2026-10-19; run date and serving revision not stated
Category: audio
Direction: Higher is better
Source type: official-leaderboard
Evaluator: Cekura Bench TTS Model Benchmarks
Status: active
Retrieved at: 2026-10-19
Methodology: Official benchmark leaderboard. Evaluator-run TTS benchmark. Cekura says each model read the same 104 agent sentences three times from one US East cloud region, using one stock voice and its own streaming protocol. Non-streaming timing runs from sending the full text to first audible sound; streaming timing runs from the last word sent to first audible sound. Reading accuracy is estimated from Deepgram Nova-3 and OpenAI Whisper transcripts, counting accepted readings without human review. Naturalness uses UTMOS, an open model predicting a listening-test score. The captured rows do not state per-cell N, CI, exact run date, serving/API revision, or exact voice ID unless shown separately.
Context: Cekura discloses that it runs the benchmark itself and sells testing and monitoring tools in the same market. The results are evaluator-published and include voice, runtime, protocol, ASR, and UTMOS effects; they are not a human-only or isolated-model evaluation.

Source Cekura Bench TTS Model Benchmarks

Microsoft-published model-inference latency (Show measurement, test conditions, and source)

Source value: 0.045
Score: 0.045
Metric: approximate model-inference latency
Unit: seconds
Benchmark version: Microsoft MAI-Voice-2.1 model page retrieved 2026-10-16
Category: audio
Direction: Lower is better
Source type: vendor-reported
Evaluator: Microsoft AI
Status: active
Retrieved at: 2026-10-16
Methodology: Vendor-reported result. The page's MAI-Voice-2.1-Flash section labels this as model-inference latency and reports approximately 45 ms. It does not specify prompt length, output length, hardware, request count, or measurement method.
Context: This is an approximate vendor-published latency figure, not a standardized quality measurement or independently reproduced benchmark.

Source Microsoft AI

Cekura TTS non-streaming latency median (Show measurement, test conditions, and source)

Source value: 289
Score: 289
Metric: Median time from complete text sent to first audible sound
Unit: ms
Benchmark version: Cekura TTS Model Benchmarks; current page snapshot 2026-10-19; run date and serving revision not stated
Category: audio
Direction: Lower is better
Source type: official-leaderboard
Evaluator: Cekura Bench TTS Model Benchmarks
Status: active
Retrieved at: 2026-10-19
Methodology: Official benchmark leaderboard. Evaluator-run TTS benchmark. Cekura says each model read the same 104 agent sentences three times from one US East cloud region, using one stock voice and its own streaming protocol. Non-streaming timing runs from sending the full text to first audible sound; streaming timing runs from the last word sent to first audible sound. Reading accuracy is estimated from Deepgram Nova-3 and OpenAI Whisper transcripts, counting accepted readings without human review. Naturalness uses UTMOS, an open model predicting a listening-test score. The captured rows do not state per-cell N, CI, exact run date, serving/API revision, or exact voice ID unless shown separately.
Context: Cekura discloses that it runs the benchmark itself and sells testing and monitoring tools in the same market. The results are evaluator-published and include voice, runtime, protocol, ASR, and UTMOS effects; they are not a human-only or isolated-model evaluation.

Source Cekura Bench TTS Model Benchmarks

Cekura UTMOS-predicted naturalness (Show measurement, test conditions, and source)

Source value: 4.32
Score: 4.32
Metric: UTMOS-predicted naturalness rating, comparison page labels scale 1 to 5
Unit: rating
Benchmark version: Cekura TTS Model Benchmarks; current page snapshot 2026-10-19; run date and serving revision not stated
Category: audio
Direction: Higher is better
Source type: official-leaderboard
Evaluator: Cekura Bench TTS Model Benchmarks
Status: active
Retrieved at: 2026-10-19
Methodology: Official benchmark leaderboard. Evaluator-run TTS benchmark. Cekura says each model read the same 104 agent sentences three times from one US East cloud region, using one stock voice and its own streaming protocol. Non-streaming timing runs from sending the full text to first audible sound; streaming timing runs from the last word sent to first audible sound. Reading accuracy is estimated from Deepgram Nova-3 and OpenAI Whisper transcripts, counting accepted readings without human review. Naturalness uses UTMOS, an open model predicting a listening-test score. The captured rows do not state per-cell N, CI, exact run date, serving/API revision, or exact voice ID unless shown separately.
Context: Cekura discloses that it runs the benchmark itself and sells testing and monitoring tools in the same market. The results are evaluator-published and include voice, runtime, protocol, ASR, and UTMOS effects; they are not a human-only or isolated-model evaluation.

Source Cekura Bench TTS Model Benchmarks

Popularity compared

How often the model is searched for, used and mentioned, compared with every model of its category.

Popularity index

0 to 100, higher is more popular

20 of 83 models

pyannote Community-173

Nemotron-3-Diarization57

Gemini 3.5 Transcribe52

Confucius4-R2T252

Granite Speech 5.0 470M TurboCTC52

Kokoro-82M51

Nova-351

Fish Audio S2 Pro51

NVIDIA Nemotron 3.5 Transcribe50

Granite Speech 5.0 470M TurboCTC NC50

Orukeet49

Lyria 3.547

Gemini 3.5 Live Translate46

Phonon-244

Aura-244

Gemini 3.1 Flash TTS44

MiniMax-Music341

Audio8-ASR-Infinite41

GPT-Realtime 2.141

MAI-Voice-2.1-Flash0

pyannote Community-1

Nemotron-3-Diarization

Gemini 3.5 Transcribe

Confucius4-R2T2

Granite Speech 5.0 470M TurboCTC

Kokoro-82M

Nova-3

Fish Audio S2 Pro

NVIDIA Nemotron 3.5 Transcribe

Granite Speech 5.0 470M TurboCTC NC

Orukeet

Lyria 3.5

Gemini 3.5 Live Translate

MAI-Voice-2.1-Flash

As of November 2026 · Source: gradually.ai popularity index

Over time

Weekly values from October 18, 2026 to November 5, 2026

MAI-Voice-2.1-FlashCartesia Ink-2Eleven v4 TurboMAI-Transcribe-2Muse Voice Transcribe

All specifications

Architecture

Model class: Low-latency speech synthesis; Source: Model class

Input and output

Input: Text; Source: Input
Output: Audio; Source: Output
Languages: 23 languages and 26 locales; Source: Languages
Latency: About 150 milliseconds end to end; Source: Latency

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; no source
Watermarking technique: Not known; no source

Licence and access

Access: Microsoft Foundry API; Source: Access

API

API model ID: mai-voice-2-1-flash; Source: API model ID

Further details

Preview terms: Public preview without an SLA; Source: Preview terms
Generation limit: Up to 45 seconds of audio; Source: Generation limit

Prices

Price: $15 per 1M charactersSource

Sources and data date

Show all 11 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.

Type	Evidence and data date
Research status	Research date October 4, 2026. 6 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication.
Not published	quality-benchmarks Original wording No exact quality/MOS benchmark; 150 ms is an operational latency claim.
Not found in the inspected sources	quality-benchmarks Note No numeric quality/MOS benchmark for exact MAI-Voice-2.1-Flash; operational latency is not a quality score.
Not found in the inspected sources	quality-benchmarks Note The exact Flash model page publishes approximate inference latency but no MOS, listener preference, or other numeric quality result; no exact Artificial Analysis row was found.
Additional source	Microsoft MAI Voice preview documentation (retrieved October 3, 2026; October 15, 2026; October 16, 2026) · Microsoft MAI Voice preview documentation · Editorial description reviewed October 3, 2026
Additional source	Microsoft MAI Voice announcement (retrieved October 3, 2026; October 4, 2026) · Microsoft MAI Voice announcement · Editorial description reviewed October 3, 2026
Additional source	Microsoft MAI-Voice-2.1 model page (retrieved October 3, 2026; October 16, 2026) · Microsoft MAI-Voice-2.1 model page · Editorial description reviewed October 3, 2026
Additional source	Cekura Bench TTS Model Benchmarks (retrieved October 19, 2026)1 2
Additional source	OpenRouter: Popularity index (derived), CC BY 4.0
Additional source	Vercel AI Gateway: Popularity index (derived), CC BY 4.0
Additional source	Ollama: cumulative downloads for the popularity index
Additional source	Replicate: cumulative runs for the popularity index
Additional source	Google Trending Now: limited feed selection for Germany and the US
Additional source	Creative Commons Attribution 4.0