Microsoft Unveils MAI-Transcribe-2-Streaming, Calls It Most Accurate Real-Time Model

Microsoft has introduced new AI speech models for real-time transcription and voice generation, offering multilingual support, low latency and tools designed for responsive conversational voice agents.

Microsoft Introduces New AI Speech Models

Microsoft has expanded its artificial intelligence portfolio with the launch of MAI-Transcribe-2-Streaming, a real-time speech-to-text model designed for fast and accurate audio transcription.

Alongside the transcription system, the company has also introduced two speech-generation models: MAI-Voice-2.1 and MAI-Voice-2.1-Flash.

The new models are aimed at developers and businesses building conversational voice agents that need quick responses, multilingual capabilities and cost-efficient AI infrastructure.

MAI-Transcribe-2-Streaming Focuses on Real-Time Accuracy

Microsoft's MAI-Transcribe-2-Streaming is designed to provide continuous speech-to-text conversion with low latency.

The model supports 60 languages and includes automatic language detection, allowing it to identify the language being spoken without requiring users to manually select it.

Microsoft AI CEO Mustafa Suleyman described the model as the world's most accurate real-time audio transcription system.

The company also said the model has achieved a leading position for transcription accuracy on the independent Artificial Analysis benchmark.

Faster Speech-to-Text Responses

A major feature of MAI-Transcribe-2-Streaming is its ability to begin generating text before a speaker has finished an entire sentence. Microsoft says the system can produce its initial text hypotheses approximately 100 milliseconds after receiving audio input.

This low-latency approach is particularly relevant for applications where conversations need to happen naturally and without long pauses between speaking and responding.

Traditional transcription systems may wait for a phrase or sentence to be completed before returning text. Microsoft's model is designed to continuously process incoming audio instead.

MAI-Voice-2.1 Brings Multilingual Speech Generation

Microsoft has also launched MAI-Voice-2.1, a multilingual text-to-speech model developed to work alongside its speech recognition technology.

The model supports 23 languages and 26 regional locales, allowing developers to create AI systems capable of communicating with users across different markets.

One of its highlighted features is cross-lingual voice identity preservation.

One AI Voice Across Multiple Languages

MAI-Voice-2.1 is designed to maintain the identity of a synthetic voice when it switches between languages.

This means a single AI-generated voice can communicate in multiple languages without automatically carrying the accent associated with one language into another.

For businesses, the technology could enable consistent virtual representatives across different countries and markets.

It could also support customer-service systems that communicate with callers in their preferred languages while maintaining a consistent voice identity.

Microsoft Launches MAI-Voice-2.1-Flash

The company has also introduced MAI-Voice-2.1-Flash, a faster version of its voice-generation model. The Flash variant is designed for applications where response speed is particularly important.

Microsoft says it retains the multilingual voice consistency capabilities of MAI-Voice-2.1 while being optimised for high-throughput enterprise workloads.

The model is intended for scenarios where AI systems need to generate speech quickly while handling a large number of interactions.

150-Millisecond Audio Generation Latency

According to Microsoft, MAI-Voice-2.1-Flash can generate 45 seconds of synthetic audio with end-to-end latency of around 150 milliseconds.

Such response times could be useful for voice-based AI agents that need to interact with users in near real time.

Reducing the delay between receiving a request and generating a spoken response is important for making automated conversations feel more natural.

Designed for Conversational Voice Agents

The three models are aimed at supporting the development of autonomous conversational voice applications. MAI-Transcribe-2-Streaming handles incoming speech, while MAI-Voice-2.1 and its Flash variant focus on producing spoken responses.

Together, these capabilities can provide developers with components needed to build AI systems that listen, understand and respond through voice.

Potential applications include customer support, virtual assistants, multilingual services and enterprise communication tools.

Focus on Speed, Language Support and Cost

Microsoft's latest speech models reflect the growing competition in AI voice technology, where developers are seeking systems that combine accuracy with low response times.

The ability to process dozens of languages can also help businesses deploy conversational AI across multiple regions without building separate voice systems for each market.

The company is positioning the models as tools that can make voice-agent development more accessible while addressing the need for fast and scalable enterprise applications.

Microsoft's Expanding AI Model Portfolio

The launch adds new capabilities to Microsoft's growing collection of internally developed AI models.

While Microsoft continues to work with external AI partners, its MAI-branded models demonstrate the company's broader push to develop proprietary artificial intelligence technologies for specific applications.

The latest speech-focused models target an area where response speed, language accuracy and natural interaction are becoming increasingly important.

Conclusion

Microsoft’s new speech AI models mark another step in the development of faster and more responsive voice-based applications. MAI-Transcribe-2-Streaming focuses on low-latency multilingual transcription, while MAI-Voice-2.1 and its Flash variant target natural, rapid speech generation. Together, the models are designed to help developers build conversational voice agents with broader language support and faster real-time interactions.