DEV.co
Open-Source LLM · microsoft

VibeVoice-1.5B

VibeVoice-1.5B is Microsoft's open-source text-to-speech model designed to generate long-form, multi-speaker conversational audio (e.g., podcasts) from text. It can synthesize up to 90 minutes of speech with up to 4 distinct speakers. The model uses continuous speech tokenizers and a diffusion-based decoding approach. It is licensed under MIT and ungated, but explicitly marked for research use only; commercial deployment is not recommended by the developers without further testing.

Source: HuggingFace — huggingface.co/microsoft/VibeVoice-1.5B
2.7B
Parameters
mit
License (OSI-approved)
Unknown
Context (tokens)
189.9k
Downloads (30d)

Key facts

Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.

FieldValue
Developermicrosoft
Parameters2.7B
Context windowUnknown
Licensemit — OSI-approved
Modality / tasktext-to-speech
Gated on HuggingFaceNo
Downloads189.9k
Likes2.4k
Last updated2026-01-22
Sourcemicrosoft/VibeVoice-1.5B

What VibeVoice-1.5B is

VibeVoice-1.5B is a transformer-based architecture combining a 1.5B-parameter Qwen2.5 LLM backbone, pre-trained acoustic and semantic tokenizers (each ~340M parameters for encoder/decoder), and a lightweight diffusion head (~123M parameters). The model operates on continuous acoustic and semantic tokens at 7.5 Hz frame rate, achieving 3200x downsampling from 24 kHz input. Training employed curriculum learning up to 65,536 token context. Inference uses Classifier-Free Guidance and DPM-Solver. The model integrates an imperceptible watermark and audible AI disclaimer in all outputs.

Quickstart

Run VibeVoice-1.5B locally

Load the open weights with 🤗 Transformers and generate — the same model, self-hosted.

quickstart.pypython
from transformers import pipelinepipe = pipeline("text-generation", model="microsoft/VibeVoice-1.5B")out = pipe("Explain retrieval-augmented generation in one sentence.",           max_new_tokens=128)print(out[0]["generated_text"])

Swap in vLLM or Ollama for production-grade serving. DEV.co can stand up the inference stack.

Deployment

How you'd run it

A typical self-hosted path — open weights, an inference server, your application.

DEV.co builds each layer — from GPU infrastructure to the application.

Best use cases

Podcast and Long-Form Audio Generation

VibeVoice's design for coherent multi-speaker dialogue and 90-minute generation capacity makes it well-suited for podcast prototyping, audiobook narration, and interview simulation in research settings.

Conversational AI Audio Research

Researchers exploring dialogue systems, narrative generation, and speaker consistency in synthetic audio can leverage the model's dialogue-aware architecture and public code without licensing friction.

Bilingual Content Prototyping (EN/ZH)

Development teams evaluating English or Mandarin Chinese audio synthesis pipelines benefit from the model's explicit support for both languages and ungated access for experimentation.

Running & fine-tuning it

ESTIMATE: The combined model (~2.7B LLM + 680M tokenizer encoders/decoders + 123M diffusion head) likely requires 8–16 GB VRAM for inference in bfloat16/fp16 (requires verification). Tokenizer inference adds secondary memory overhead. Exact precision and batch-size behavior not specified in card; benchmarks unavailable.

The model card states that 'pre-trained tokenizers are frozen; only the LLM and diffusion head parameters are trained' during base training. Fine-tuning feasibility (LoRA, QLoRA, full) is not discussed. Given the frozen tokenizers, any fine-tuning would be limited to the LLM and diffusion head layers. Requires review of training infrastructure and gradient checkpointing support.

When to avoid it — and what to weigh

  • Production/Commercial Voice Cloning without Consent — The model explicitly prohibits voice impersonation without recorded consent. Do not use to clone real individuals' voices for advertising, social engineering, authentication, or satire.
  • Real-Time or Low-Latency Voice Conversion — VibeVoice is not designed for live telephone or video-conference applications. The model targets batch processing of long-form content, not real-time inference.
  • Non-English/Non-Chinese Languages — The model is trained only on English and Mandarin Chinese data. Outputs in other languages are unsupported and may produce unintelligible or offensive content.
  • Disinformation or Deepfake Audio Deployment — The model's responsible-use guidelines explicitly prohibit creating audio presented as genuine recordings of real people or events without disclosure. Watermarking and logging are present but do not eliminate misuse risk.

License & commercial use

MIT License. This is a permissive OSI-approved open-source license allowing modification, distribution, and commercial use subject to including the license and copyright notice. The license itself does not restrict use.

CAUTION: Although the MIT license permits commercial use, the model card explicitly states: 'We do not recommend using VibeVoice in commercial or real-world applications without further testing and development. This model is intended for research and development purposes only.' Additionally, the model prohibits specific uses (voice impersonation, disinformation, real-time voice conversion, non-consented speaker cloning). Any commercial deployment must (1) respect these explicit restrictions, (2) verify compliance with applicable laws and regulations, and (3) obtain independent security/bias testing. Commercial use requires legal and technical review before production.

DEV.co evaluation signals

Editorial assessment — not user reviews. Directional, with an explicit confidence level.

SignalAssessment
MaintenanceActive
DocumentationAdequate
License clarityClear
Deployment complexityModerate
DEV.co fitGood
Assessment confidenceHigh
Security considerations

The model includes two protective measures: (1) imperceptible watermark embedded in generated audio for provenance verification, and (2) audible AI-generated disclaimer automatically inserted into all outputs. Hashed inference logs are recorded for abuse pattern detection. However, these mitigations do not eliminate misuse risk (watermarks can be bypassed, disclaimers can be edited, logs are reactive). The model inherits biases and errors from its Qwen2.5 base LLM. High-quality synthetic speech poses inherent deepfake and disinformation risks. Users are responsible for lawful deployment and data sourcing.

Alternatives to consider

Qwen2-Audio (Qwen team)

Multi-modal audio-language model offering different architectural approach; may provide better unsupervised audiovisual understanding, but less specialized for conversational TTS.

MetaVoice-1B (MetaVoice team)

Lightweight open-source TTS focused on naturalness and few-shot speaker adaptation; smaller footprint but designed for shorter utterances, not long-form dialogue.

Parler TTS (OpenVoiceOS)

AGPL-licensed lightweight TTS for local deployment; no multi-speaker support and does not target long-form or dialogue-aware synthesis, but avoids commercial restrictions in model intent.

Software development agency

Ship VibeVoice-1.5B with senior software developers

Download from Hugging Face (microsoft/VibeVoice-1.5B), review the GitHub repository, and consult the technical report. For commercial deployment or fine-tuning, contact Microsoft Research at [email protected]. Conduct independent testing and compliance review before production use.

Talk to DEV.co

Related open-source tools

Surfaced by semantic similarity across the DEV.co open-source index.

Related on DEV.co

Explore the category and the services that help you build with it.

VibeVoice-1.5B FAQ

Can I use VibeVoice commercially to build a podcast generation or audiobook service?
The MIT license technically permits commercial use, but Microsoft explicitly recommends against commercial deployment without further testing and development. Any commercial use must comply with all stated out-of-scope restrictions (no voice cloning without consent, no disinformation, etc.) and obtain independent security and legal review. This is not a straightforward yes.
What are the GPU memory requirements for inference?
Not clearly stated. The model card provides parameter counts (~2.7B LLM + tokenizer/diffusion components) but no VRAM benchmarks. Estimate 8–16 GB in bfloat16/fp16 for inference; requires empirical testing with your hardware and batch size.
Does VibeVoice support real-time streaming or low-latency synthesis?
No. VibeVoice is designed for batch long-form audio generation (up to 90 minutes per call). A separate streaming variant (VibeVoice-Realtime-0.5B) exists but is not the subject of this release. Do not use for telephony, live conferencing, or interactive audio applications.
Can I fine-tune VibeVoice on my own data?
Not directly documented. The model card indicates that acoustic and semantic tokenizers are frozen during training, with fine-tuning limited to the LLM and diffusion head. Fine-tuning code, datasets, and tutorials are not provided in the card. Requires review of the GitHub repository and direct contact with Microsoft Research.

Work with a software development agency

From first prototype to production, DEV.co delivers software development services around tools like VibeVoice-1.5B. Our software development agency staffs experienced software developers and web developers for custom software development, web development, integrations, and ongoing support across open-source llms and beyond.

Get Started with VibeVoice-1.5B

Download from Hugging Face (microsoft/VibeVoice-1.5B), review the GitHub repository, and consult the technical report. For commercial deployment or fine-tuning, contact Microsoft Research at [email protected]. Conduct independent testing and compliance review before production use.