Microsoft launches MAI-Transcribe-2-Streaming audio

The new model family combines an independently benchmarked real-time transcription engine with two low-latency speech synthesis models for conversational agents.

On October 1, 2026, Microsoft AI released three proprietary audio models designed for voice agents: MAI-Transcribe-2-Streaming for live speech recognition, alongside MAI-Voice-2.1 and MAI-Voice-2.1-Flash for speech synthesis. All three models are available through the MAI Playground, Microsoft Foundry, Vercel AI Gateway, and OpenRouter.

MAI-Transcribe-2-Streaming tops speech benchmarks

MAI-Transcribe-2-Streaming handles 60 languages with continuous, automated spoken-language detection during incoming audio streams. In an independent evaluation conducted by Artificial Analysis across 38 competing speech models, the system took first place overall, posting a 2.5% Word Error Rate (WER) across both partial and final outputs.

For engineers building conversational agents, latency is the defining metric. The model outputs its first partial hypotheses roughly 100 milliseconds after receiving audio, registering latencies of 0.12 seconds for partial transcripts and 0.13 seconds for final outputs. That early window lets applications trigger database queries or tool calls before the caller finishes speaking. Microsoft introduced a promotional launch rate of $0.54 per audio hour ($9.00 per 1,000 minutes) through December 2026, leaving standard pricing for 2027 unannounced.

Multilingual voices with consistent speaker identity

On the speech generation side, Microsoft introduced MAI-Voice-2.1 and MAI-Voice-2.1-Flash across 23 languages and 26 locales. The core technical feature is cross-lingual voice consistency: a single speaker profile preserves its recognizable vocal identity across different languages while adopting the native accent and pronunciation of each locale.

The two synthesis models serve distinct workloads. The Flash edition prioritizes quick turnaround, running with an end-to-end latency of 150 milliseconds at a cost of $15 per million characters. Microsoft claims inference is 55% faster than comparable industry models, though the company did not specify which baselines were tested. The standard MAI-Voice-2.1 targets long-form speech such as audiobooks and recorded presentations, priced at $22 per million characters.

Both models support instant voice cloning from audio samples between 5 and 60 seconds, subject to formal consent verification controls. Microsoft also reported results from an internal blind test with 4,000 listeners, in which 50.3% of respondents judged the synthesized voice as human or more human than real recordings. Because this metric originates from an internal evaluation by the vendor, it lacks independent third-party confirmation.

The fight for the millisecond budget

Conversational systems need to respond in under 700 milliseconds to avoid awkward lags. When transcription takes 100 milliseconds and synthesis takes 150 milliseconds, developers gain essential headroom for the underlying language model to reason without stalling the conversation. With this release, Microsoft is competing directly against voice specialists such as Deepgram, ElevenLabs, and Cartesia.

Despite solid benchmark figures, teams moving these models into production face practical hurdles. In the Azure Foundry catalog, the voice models remain in public preview, which carries no service level agreement (SLA). In addition, direct integration with the LiveKit real-time audio framework has not received a formal release date.

Sources

• Microsoft AI: https://microsoft.ai/news/our-first-streaming-transcription-model/
• Microsoft Learn: https://learn.microsoft.com/en-us/azure/ai-services/speech-service/mai-voices
• Artificial Analysis: https://artificialanalysis.ai/speech-to-text/streaming
• Marktechpost: https://www.marktechpost.com/2026/10/02/microsoft-ai-releases-mai-transcribe-2-streaming-1-real-time-speech-to-text-model-on-artificial-analysis/
• TestingCatalog: https://x.com/testingcatalog/status/2105723322444959922