Transcribe long meetings and synthesize multi-speaker podcasts with open-source voice AI
Feed it a 60-minute meeting recording and it extracts who said what and when in structured text; feed it text and it synthesizes conversations up to 90 minutes long. Microsoft has released VibeVoice, a collection of models covering both speech recognition and synthesis. Its defining feature is processing long audio in one pass to maintain context, rather than chopping it into small pieces.
Recent releases: Streaming recognition and CPU-only engine
On September 3, 2026, VibeVoice-ASR-Streaming was released, which continuously transcribes who is saying what as audio arrives. This model supports custom hotwords and 10 languages. Earlier, on July 23, VibeVoice-ASR-BitNet was released, enabling real-time inference on edge CPUs. By applying heterogeneous quantization (I8_S + I2_S), the model size was reduced from 4.62GB to 1.58GB, and it performs real-time processing (RTF < 1) on three or more CPU threads without a GPU.
Recognition model that reads 60 minutes of audio at once
VibeVoice-ASR processes audio up to 60 minutes long in a single pass within a 64K token length. Unlike existing models that cut audio into short fragments and lose global context, this model maintains speaker tracking and semantic consistency throughout the hour. The output is a structured transcript containing the speaker (Who), timestamp (When), and content (What). Providing specific names, technical terms, or background information as hotwords improves recognition accuracy in that domain. It also supports a streaming variant that outputs text while the audio is still arriving.
Long conversation synthesis and real-time TTS
VibeVoice-TTS synthesizes conversational audio up to 90 minutes long in one go, supporting up to four different speakers to maintain natural turn-taking and speaker consistency. It supports multiple languages, including English and Chinese, and is based on research accepted as an Oral at ICLR 2026. Currently, VibeVoice-Realtime-0.5B handles real-time text-to-speech conversion. This model is 0.5B parameters in size, making it favorable for deployment, and shows a latency of approximately 300 milliseconds to the first sound. It supports streaming text input and stable generation of long audio up to about 10 minutes. On December 16, 2025, multilingual speakers for 9 languages and 11 English style speakers were added for experimental use.
Limitations and cautions as a research tool
This model is intended for research and development purposes only, and commercial or real-world application is not recommended. Additional testing and development are required. It may contain biases, errors, and omissions inherited from the base model (Qwen2.5 1.5b), and there is a risk that high-quality synthesized audio could be misused for impersonation, fraud, or the spread of misinformation. You must verify the reliability and accuracy of generated content, and it is recommended to explicitly state that it is AI-generated. Legal compliance and responsible use are prerequisites.
