VibeVoice-1.5B

VibeVoice-1.5B

Turns written conversations into spoken audio with up to four speakers and about 90 minutes of speech, an open Microsoft text-to-speech model for research on podcast-style dialogue in English and Chinese.

Made by microsoft on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: VibeVoice-1.5B is a text-to-speech model from Microsoft for long conversations with several voices, such as a podcast. It can produce about 90 minutes of speech with up to 4 speakers. Where many speech models stop at one or two voices, this one keeps speakers consistent and handles turn-taking. It supports English and Chinese.

Who it suits: It suits speech and audio researchers who want to explore realistic dialogue generation, since the card limits its use to research.

Title: Who uses this model
Industry | Who uses it | How they use it | Example
Speech research | Researchers | Study long, multi-speaker audio generation | Generate a four-person conversation to examine speaker consistency
Media research | Audio engineers | Prototype podcast-style audio for study | Render a written interview as a two-voice recording
Language technology | Developers | Test English and Chinese dialogue speech | Compare output in the two supported languages

If you want music, sound effects or speech in other languages, you can skip it.

Where to use it:

• Python on Windows, macOS or Linux
• Research setups that load models with the transformers library

Title: Jobs it does
Use case | Who | What they get back
Podcast-style dialogue | Researchers | Spoken audio for a written conversation with several voices
Long-form speech | Audio engineers | Up to about 90 minutes of continuous speech
Two-language speech | Developers | English or Chinese speech from text

How to set it up: You need Python and about 5.4 GB of disk space for the BF16 weights. The card does not list install commands and sends you to the project's GitHub README.

• Python with pip
• The huggingface_hub package
• The install steps from the project's GitHub README

1. Download the model files with the two Hugging Face commands.
2. Open the project's GitHub README and follow its installation section.
3. Run the model as that README describes.

pip install -U huggingface_hub
hf download microsoft/VibeVoice-1.5B

How to use it: The card gives no run command and points to the GitHub README for usage. In general you provide a written conversation and the model returns audio in which each speaker has a consistent voice. Use only voices you have the right to use.

Example: You supply a short script for a two-person interview, with each line marked for its speaker, and run the model. The output is spoken audio that follows the dialogue, with each speaker's voice staying steady across turns.

For newcomers and experts: Text-to-speech software reads text aloud. This model is built for long talks between several people, like a podcast episode. A first small test is a short two-speaker script from the project README.

A Qwen2.5-1.5B language model reads the text and the flow of dialogue. A diffusion head with 4 layers and about 123M parameters then predicts the acoustic detail. Two speech tokenizers run at 7.5 Hz, which keeps long sequences affordable. The tokenizers were pre-trained and frozen, and then only the language model and the head were trained, with context growing up to 64K tokens.

Written dialogue goes in
Qwen2.5-1.5B reads context and flow
Diffusion head predicts acoustic detail
Acoustic tokenizer decodes to speech

Good to know: The card names other sizes. The Large version shows about 45 minutes of generation and its weights are listed as disabled. Other people have published 9 quantized and 17 fine-tuned versions. The card forbids voice impersonation without recorded consent, fake recordings of real people and live voice conversion.

Title: Generation length by model
VibeVoice-1.5B: 90 min
VibeVoice-Large: 45 min

• License: MIT.
• Cost to run: none named.
• Data it reaches: the files come from Hugging Face, and the model then runs on your own computer.
• Not for: anything beyond research, other languages, music or sound effects, impersonating real people, or live deepfake calls.

On Hugging Face, 14 November 2026

• Task: text to speech
• Library: transformers
• Parameters: 2.7B
• Weights: about 5.4 GB in BF16; about 1.4 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 217,923
• Likes: 2,520
• License: mit
• Languages: en, zh
• Quantized versions: 9
• Fine-tunes: 18
• Tags: transformers, safetensors, vibevoice, text-to-audio, podcast, text-to-speech

Links

• Model: https://huggingface.co/microsoft/VibeVoice-1.5B
• Files: https://huggingface.co/microsoft/VibeVoice-1.5B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download microsoft/VibeVoice-1.5B

Then follow the usage in its model card. Nothing in this kit runs by itself.