VibeVoice:微软开源的前沿语音 AI 家族,ASR 一次跑60分钟,CPU 也能实时推理
VibeVoice: Microsoft's Open-Source Frontier Voice AI — 60-Minute ASR in One Pass, Real-Time on CPU
三个模型,覆盖语音 AI 的三个核心场景
微软 2025 年 8 月开始陆续开源 VibeVoice 系列——一套前沿语音 AI 模型家族,覆盖长文本转语音、长音频识别和边端实时推理三个核心场景。
⭐ 53,474,MIT 许可,持续更新到 2026 年 9 月。以下按最新进展从大到小介绍。
最新:ASR-BitNet——无 GPU,三线程,实时
2026年7月23日发布的 VibeVoice-ASR-BitNet 是目前最值得关注的更新:
通过异构量化(I8_S + I2_S),模型从 4.62 GB 压缩到 1.58 GB,在 3 个以上 CPU 线程上实现 RTF < 1 的实时推理——不需要 GPU。
这意味着完整的长音频识别能力(60 分钟单次处理、说话人识别、时间戳、50+ 语言)可以在普通笔记本 CPU 上跑起来,不需要 CUDA 环境。
ASR-7B:一次处理60分钟,说清楚”谁/什么时候/说了什么”
VibeVoice-ASR(7B 参数)解决的是传统 ASR 的一个核心限制:大多数识别模型需要把长音频切成短片段分批处理,丢失跨段的全局上下文,说话人追踪更是挑战。
VibeVoice-ASR 直接接受最多 60 分钟连续音频输入(64K token 长度),单次完成三件事:
| 能力 | 说明 |
|---|---|
| Who(说话人识别/区分) | 在整段音频里追踪和区分多个说话人 |
| When(时间戳) | 每个片段精确定位到对应时间 |
| What(内容转录) | 50+ 语言多语言识别 |
还支持用户提供自定义热词(专有名词、技术术语、背景信息),提升领域识别准确率。
集成路径:
- HuggingFace Transformers 直接使用(2026-03-06 加入)
- vLLM 加速推理(已支持)
- Azure AI Foundry Labs(2026-03-12 上线)
from transformers import pipeline
asr = pipeline("automatic-speech-recognition", model="microsoft/VibeVoice-ASR")
result = asr("your_audio.wav")
# 返回结构化输出:说话人 + 时间戳 + 文字
Realtime-0.5B:300ms 首字延迟的流式 TTS
VibeVoice-Realtime-0.5B 是轻量实时 TTS 模型,定位和 ASR-7B 互补——一个认声音,一个合成声音:
- 首字延迟:约 300 毫秒(文字输入到开始出声)
- 参数量:0.5B,面向部署友好
- 流式输入:支持边写边合成,不需要等完整文本
- 长文本:单次约 10 分钟,超出可流式接续
- 多语言:9 种语言实验性支持(德、法、意、日、韩、荷、波、葡、西)+ 11 种英语风格声线
2025 年 12 月 16 日更新加入了多语言声音和不同英语风格(新闻播报、讲故事、技术讲解等),可以直接在 Colab 试用。
TTS-1.5B:已下线,原因是滥用
项目的起点是 2025 年 8 月发布的 VibeVoice-TTS-1.5B——一个能在单次推理里合成最多 90 分钟、最多 4 个说话人的长文本多说话人 TTS 模型,并以 ICLR 2026 Oral 被接收。
但开源后不久(2025-09-05),微软发现有人以与初衷不符的方式使用这个模型——代码随即从仓库中下线,只保留论文和技术文档。权重仍在 HuggingFace 上,但代码不再提供。
这是语音 AI 开源里一个值得记录的案例:高质量合成声音的滥用风险促使了一次明确的撤库决定。
核心技术:7.5 Hz 连续声学标记 + 下一个 Token 扩散
VibeVoice 系列的共同技术基础:
连续语音 Tokenizer(声学 + 语义)在 7.5 Hz 超低帧率下运行——比传统方法低得多——同时保留音频保真度,大幅减少长序列的计算开销。这是处理 60/90 分钟超长音频的关键。
下一个 Token 扩散框架:
- LLM 理解文本上下文和对话流(基于 Qwen2.5)
- 扩散头生成高保真声学细节
这一架构把 LLM 的语义理解能力和扩散模型的高质量声学生成结合在同一个框架里。
模型概览
| 模型 | 参数 | 用途 | 状态 |
|---|---|---|---|
| VibeVoice-ASR | 7B | 长音频识别(60分钟,多说话人,时间戳) | ✅ 可用 |
| VibeVoice-ASR-BitNet | 1.58 GB | 边端 CPU 实时推理 | ✅ 可用 |
| VibeVoice-Realtime | 0.5B | 实时流式 TTS(300ms 延迟) | ✅ 可用 |
| VibeVoice-TTS | 1.5B | 超长多说话人 TTS(90分钟) | ⚠️ 权重保留,代码下线 |
快速上手
ASR(标准 GPU):
pip install transformers
from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor
# 详见 docs/vibevoice-asr.md
ASR-BitNet(CPU 无 GPU):
git clone https://github.com/microsoft/VibeASR.cpp
# 详见 README,需要 3+ CPU 线程
Realtime TTS(Colab 一键运行):
直接打开 Colab 链接 即可试用。
总结
VibeVoice 在语音 AI 里做了三件有实质区分度的事:
- ASR 把”几分钟短音频”升级到”60分钟完整会议/演讲”,同时输出说话人、时间戳、内容三合一;
- ASR-BitNet 把这个能力压进 CPU,让没有 GPU 的环境也能实时跑;
- Realtime TTS 提供 300ms 延迟的流式合成,适合对话和实时场景。
53k Star,微软研究院出品,持续更新。对于需要在生产里处理长音频、或者需要 CPU 部署的语音场景,VibeVoice 是目前开源里覆盖最完整的选项之一。
GitHub: microsoft/VibeVoice ⭐53474
项目主页: microsoft.github.io/VibeVoice
ASR Playground: aka.ms/vibevoice-asr
VibeVoice: Microsoft’s Open-Source Frontier Voice AI Family
Microsoft has been open-sourcing VibeVoice since August 2025 — a family of frontier Voice AI models covering long-form TTS, long-audio recognition, and edge real-time inference.
⭐53,474, MIT license, actively updated through September 2026. Three models, three use cases.
Latest: ASR-BitNet — Real-Time on CPU, No GPU
The most significant recent update (July 23, 2026) is VibeVoice-ASR-BitNet:
Via heterogeneous quantization (I8_S + I2_S), the model is compressed from 4.62 GB to 1.58 GB, achieving real-time inference (RTF < 1) on 3+ CPU threads — no GPU required.
The full 60-minute single-pass ASR capability (speaker diarization, timestamps, 50+ languages) now runs on an ordinary laptop CPU without CUDA.
- Code: microsoft/VibeASR.cpp
- Model: HuggingFace/VibeVoice-ASR-BitNet
- Paper: arxiv.org/abs/2607.21075
ASR-7B: 60 Minutes in One Pass — Who, When, What
VibeVoice-ASR (7B) addresses a core limitation of traditional ASR: most models slice long audio into short chunks (losing global context) and struggle with multi-speaker tracking across segments.
VibeVoice-ASR accepts up to 60 minutes of continuous audio (64K token length) in a single pass and jointly produces:
| Output | Description |
|---|---|
| Who | Speaker diarization — tracking and distinguishing speakers across the full recording |
| When | Timestamps for each segment |
| What | Transcription in 50+ languages |
Plus customized hotwords: provide specific names, technical terms, or background info to guide recognition on domain-specific content.
Integration paths:
- HuggingFace Transformers (added March 2026)
- vLLM for accelerated inference
- Azure AI Foundry Labs
from transformers import pipeline
asr = pipeline("automatic-speech-recognition", model="microsoft/VibeVoice-ASR")
result = asr("your_audio.wav")
# Returns structured output: speaker + timestamp + text
Realtime-0.5B: 300ms First-Audio Streaming TTS
VibeVoice-Realtime-0.5B is the lightweight real-time TTS counterpart:
- ~300ms first audible latency from text input to audio start
- 0.5B parameters — deployment-friendly
- Streaming text input — synthesize as you type, no wait for complete text
- Long-form: ~10 minutes single pass, streamable for longer
- Multilingual: 9 experimental languages (DE, FR, IT, JP, KR, NL, PL, PT, ES) + 11 English style voices
Try it in Colab.
TTS-1.5B: Code Removed Due to Misuse
The series started with VibeVoice-TTS-1.5B in August 2025 — a long-form multi-speaker TTS capable of generating up to 90 minutes with up to 4 distinct speakers in a single inference pass, accepted as an ICLR 2026 Oral.
Shortly after release (September 5, 2025), Microsoft discovered uses inconsistent with the stated research intent. The code was removed from the repository. The weights remain on HuggingFace; the code does not. A notable case study in open-source speech AI: high-quality voice synthesis can be misused, and Microsoft drew a clear line.
Core Technology: 7.5 Hz Continuous Tokenization + Next-Token Diffusion
Common technical foundation across the VibeVoice family:
Continuous speech tokenizers (Acoustic + Semantic) operating at an ultra-low 7.5 Hz frame rate — far lower than conventional approaches — while preserving audio fidelity and dramatically reducing compute for long sequences. This is what makes 60-90 minute single-pass generation feasible.
Next-token diffusion framework: an LLM (Qwen2.5) handles textual context and dialogue flow; a diffusion head generates high-fidelity acoustic details. LLM semantic understanding + diffusion acoustic quality in one architecture.
Model Summary
| Model | Size | Use | Status |
|---|---|---|---|
| VibeVoice-ASR | 7B | Long-audio recognition (60 min, multi-speaker, timestamps) | ✅ Available |
| VibeVoice-ASR-BitNet | 1.58 GB | Edge CPU real-time inference | ✅ Available |
| VibeVoice-Realtime | 0.5B | Real-time streaming TTS (300ms latency) | ✅ Available |
| VibeVoice-TTS | 1.5B | Ultra-long multi-speaker TTS (90 min) | ⚠️ Weights only, code removed |
Summary
VibeVoice makes three meaningfully differentiated contributions to open-source Voice AI:
- ASR upgrades from “short audio clips” to “60-minute complete meetings/talks,” jointly outputting speaker, timestamp, and content
- ASR-BitNet compresses this capability to CPU, enabling real-time inference without GPU hardware
- Realtime TTS delivers 300ms latency streaming synthesis for conversational and live use cases
53k stars, Microsoft Research, ongoing updates. For production workloads requiring long-audio processing or CPU-only deployment, VibeVoice is one of the most complete open-source options available.
GitHub: microsoft/VibeVoice ⭐53474
Project Page: microsoft.github.io/VibeVoice
ASR Playground: aka.ms/vibevoice-asr
关于本站 · 免责声明
🍄 Mushroom Research Blog 是非营利、免费公开的个人科技观察博客与公众号 XStack18,不接受商业合作、不代表任何企业或机构立场,也不谋求商业利益。我们以个人视角客观中立地记录和分析 AI、Web3 等领域的最新模型发布与技术动态——不止转述新闻标题或二手信息,而是给出有独立思考的深入分析,希望帮更多人获得有价值的一手科技认知。
⚠️ 文中介绍的开源代码与模型,仅供学习交流与技术借鉴。它们大多仍处于早期阶段,有待进一步研究和验证,请勿直接用于工作或生产环境;如需采用,请先自行充分测试,并核实其许可证与安全性。
Open-source code and models featured here are shared for learning and reference only. Most are early-stage and still need further study and verification — please don't use them directly in your work or in production. Test them thoroughly and check their licenses and security first.
- 本站文章均为作者基于公开信息的个人研究与观点整理,不代表文中提及的任何公司、产品、模型的官方立场,未与其构成商业关联或合作关系。
- 科技行业信息更新极快,我们尽力保证内容准确、及时,但不对完整性、实时性做绝对保证,具体请以相关企业/项目官方公告为准。
- 文中引用的第三方商标、产品名称、图片、数据等版权归原权利人所有,我们会尽量注明来源;如你认为存在版权疑问或侵权,请通过下方邮箱联系我们,收到通知后会尽快核实处理(更正、加注来源或删除)。
- 文章内容仅为技术科普与个人观点,不构成投资、法律或其他专业建议,据此进行任何决策的后果需自行判断和承担。
📮 侵权 / 勘误 / 合作咨询:hello@mushroom.cv
💬 评论与讨论
使用 GitHub 账号登录后发表评论