NVIDIA VoiceChat 11B:第一个支持工具调用的开源全双工语音 Agent
NVIDIA VoiceChat 11B: The First Open-Source Full-Duplex Speech Agent with Tool Calling
by Mycelium Protocol
语音 AI 的主流架构是三段级联:ASR(语音转文字)→ LLM(推理)→ TTS(文字转语音)。这个方案的代价是延迟叠加、状态割裂、真实打断(barge-in)难以实现。
NVIDIA NemotronLabs VoiceChat 11B 用一个统一模型做完了这三件事——同时还实现了一个此前无人做过的功能:在对话进行中触发工具调用,工具执行时模型自动说一句占位语音(on-hold message),工具返回后无缝继续,整个过程保持自然的对话流。
HuggingFace: https://huggingface.co/nvidia/NVIDIA-NemotronLabs-VoiceChat-11B | ⭐ 239 likes
GitHub: https://github.com/NVIDIA-NeMo/Speech/tree/nemotron-labs-voicechat
发布日期:2026-08-03 | License: OpenMDW 1.1(研究用途)
核心指标
| 维度 | 数值 |
|---|---|
| 参数量 | 11B |
| 转轮延迟 | ~450 ms |
| VoiceBench 排名(开源全双工) | #2 |
| 工具调用 | ✅ 开源全双工首个 |
| 架构类型 | Hybrid Mamba/Transformer |
架构:四个模块,一次前向传播
用户语音 (16kHz)
│
▼
Fast Conformer 语音编码器
(Nemotron-Speech-Streaming-En-0.6b)
│
▼
Nemotron Nano v2 9B LLM 主干
(Hybrid Mamba/Transformer)
│
├──────────────────────────────────┐
▼ ▼
NVIDIA TTS 解码器 + 编解码器 工具调用通道(独立输出通道)
Agent 语音 (22.05kHz) 工具调用脚本
- Fast Conformer:流式语音编码,把原始音频映射为音频 token
- Nemotron Nano v2 9B:LLM 主干,预测文本 token
- TTS 解码器:把文本 token 转为语音 codec,实时合成
- 独立工具调用通道:和语音流并行输出,不打断语音生成
传统 ASR→LLM→TTS 三跳延迟约 1-2 秒,这套统一架构做到 ~450ms。
工具调用:边说话边调工具
这是 VoiceChat 最关键的突破。
传统语音助手触发工具时,通常会有一段”请稍等”的停顿,然后说”我帮你查一下……”——期间对话流被打断,用户体验差。
VoiceChat 的实现方式:
- LLM 生成触发工具调用的文本时,独立工具调用通道立即发出工具调用脚本
- 与此同时,TTS 通道说出为这个工具预定义的 on-hold 占位语音(例如”让我查一下……”)
- 工具返回结果后,模型无缝衔接继续对话
整个过程从用户角度看是连续的自然对话,没有沉默停顿,没有体验断层。
全双工能力
真实打断(Barge-in):用户说话时可以直接打断模型,模型立即停止输出并响应。不是”等我说完”,而是真正并发的双向流。
自然转轮(Turn-taking):基于 RNNT 的转轮检测,语音结束时自动触发响应,约 450ms 延迟,接近真实人类对话节奏。
实时用户转写:对话过程中同步输出用户语音的文字转写,可用于日志和后处理。
部署方式
离线推理(HuggingFace checkpoint):
# 加载 HF checkpoint 做批量语音转语音测试
# 详见 GitHub 仓库 nemotron-labs-voicechat 分支
实时交互流(NVIDIA NIM 容器):
# 官方 NIM 容器(amd64,需要 NVIDIA A100/H100/H200/B100/B200/RTX-6000)
docker pull nvcr.io/nim/nvidia/nemotron-labs-voicechat
支持的硬件:NVIDIA A100 / H100 / H200 / B100 / B200 / RTX-6000(Linux)
官方 NIM 容器仅支持 amd64。在 NVIDIA DGX Spark(aarch64/GB10)上运行需要自行重建服务栈——jxlarrea/nvidia-voicechat-spark 仓库记录了一次完整的 DGX Spark 移植实验,结论是最优配置 RTF ~1.13,尚未达到实时(<1.0),供研究参考。
技术背景
VoiceChat 11B 基于以下组件:
- 基底模型:NVIDIA Nemotron Nano 9B v2
- 语音编码器:Nemotron-Speech-Streaming-En-0.6b(Fast Conformer)
- 训练数据:~550k 小时音频,含真实语音(Fisher/LibriVox/LibriTTS)+ 合成语音 + Nemotron 5.5 文本数据
- 相关论文:arXiv 2410.17196 / 2503.04721 / 2604.04847 / 2505.15670 / 2507.08128
注意事项
- License 为 OpenMDW 1.1,仅限研究用途,不适合商业部署
- 目前为 v1 版本,NVIDIA 定性为”研究就绪”,非生产就绪
- RNNT 转轮检测对非语音瞬态(键盘声等)有误触发
- EarTTS 语音质量当前为研究级,和商业 TTS 有差距
Mycelium Protocol — 追踪 AI 系统的底层演化
关于 Mycelium
菌丝协议。持续追踪 AI 工具、系统和实验的内容节点。
NVIDIA VoiceChat 11B: The First Open Full-Duplex Voice Agent with Tool Calling
by Mycelium Protocol
The mainstream architecture for voice AI is a three-stage cascade: ASR (speech-to-text) → LLM (reasoning) → TTS (text-to-speech). The cost: stacked latency, split state, and true barge-in is hard to implement.
NVIDIA NemotronLabs VoiceChat 11B does all three in a single unified model — and adds a capability no open model has offered before: tool calling mid-conversation. When a tool is triggered, the model speaks a predefined on-hold message while the tool executes, then seamlessly continues — the conversation never stops.
HuggingFace: https://huggingface.co/nvidia/NVIDIA-NemotronLabs-VoiceChat-11B
GitHub: https://github.com/NVIDIA-NeMo/Speech/tree/nemotron-labs-voicechat
Released: 2026-08-03 | License: OpenMDW 1.1 (research only)
Core Metrics
| Dimension | Value |
|---|---|
| Parameters | 11B |
| Turn-taking latency | ~450 ms |
| VoiceBench (open FD models) | #2 |
| Tool calling | ✅ First open FD model |
| Architecture | Hybrid Mamba/Transformer |
Architecture: Four Modules, One Forward Pass
User audio (16kHz)
│
▼
Fast Conformer Speech Encoder
(Nemotron-Speech-Streaming-En-0.6b)
│
▼
Nemotron Nano v2 9B LLM backbone
(Hybrid Mamba/Transformer)
│
├──────────────────────────────────┐
▼ ▼
NVIDIA TTS Decoder + Codec Tool-calling channel (separate output)
Agent speech (22.05kHz) Tool-calling scripts
- Fast Conformer: streaming speech encoding, maps raw audio to audio tokens
- Nemotron Nano v2 9B: LLM backbone, predicts text tokens
- TTS decoder: converts text tokens to speech codec, real-time synthesis
- Separate tool-calling channel: runs parallel to speech output, no interruption
Traditional ASR→LLM→TTS cascades accumulate ~1-2s of latency. This unified architecture achieves ~450ms.
Tool Calling: Tools Fire While the Model Speaks
This is VoiceChat’s key breakthrough.
Traditional voice assistants pause when a tool is triggered — a “please wait” silence, then “let me check…” — breaking the conversational flow.
VoiceChat’s approach:
- When the LLM generates text that triggers a tool call, the independent tool-calling channel emits the tool script immediately
- Simultaneously, the TTS channel speaks the tool’s predefined on-hold message (“Let me look that up…”)
- When the tool returns, the model continues the conversation seamlessly
From the user’s perspective: no silence, no broken flow — a continuous, natural conversation.
Full-Duplex Capabilities
True barge-in: the user can interrupt at any time; the model immediately stops and responds. Not “wait for me to finish” — genuinely concurrent bidirectional streams.
Natural turn-taking: RNNT-based turn detection, ~450ms response latency, close to natural human conversation rhythm.
Live user transcription: real-time text output of user speech alongside the conversation, usable for logs and post-processing.
Deployment
Offline inference (HuggingFace checkpoint): batch speech-to-speech testing.
Interactive streaming (NVIDIA NIM container):
docker pull nvcr.io/nim/nvidia/nemotron-labs-voicechat
Supported hardware: NVIDIA A100 / H100 / H200 / B100 / B200 / RTX-6000 (Linux, amd64).
The official NIM container is amd64-only. Running on NVIDIA DGX Spark (aarch64/GB10) requires rebuilding the serving stack — the jxlarrea/nvidia-voicechat-spark repository documents a full Spark porting experiment, concluding that the best configuration reaches RTF ~1.13 (real-time requires < 1.0), for research reference.
Notes
- License: OpenMDW 1.1 — research use only, not suitable for commercial deployment
- v1 is “research-ready,” not production-ready per NVIDIA’s own characterization
- RNNT turn detection can hallucinate from non-speech transients (keyboard clicks, etc.)
- EarTTS voice quality is currently research-grade, below commercial TTS
Mycelium Protocol — tracking the deep evolution of AI systems
© 2026 Mycelium Protocol. All rights reserved.
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:[email protected]
💬 评论与讨论
使用 GitHub 账号登录后发表评论