Releases·Model Release
NemotronLabs VoiceChat: An Open Full-duplex Speech-to-Speech Model with Tool Calling Capabilities
NemotronLabs introduced VoiceChat, an open full-duplex speech-to-speech model with native tool calling, in an arXiv paper. It unifies listening, transcribing, reasoning, tool invocation, and speaking in one streaming architecture while preserving natural conversational timing. The model posts strong interruption-handling and tool-selection results on Full-Duplex-Bench, though argument accuracy and end-to-end execution remain weak spots.
0 engagement1 sourceFri, Sep 18, 2026, 04:27 PM
AI Overview
- NemotronLabs VoiceChat pairs a streaming speech encoder with a decoder-only LM, parallel text/function-call streams, RNN-T transcription, and streaming TTS in one architecture.
- On Full-Duplex-Bench 1.0 it posts 100% takeover after user interruptions, a 4.33/5 post-interruption quality score, and 93% backchannel resume rate on FDB 1.5.
- It scores 55.1 normalized on VoiceBench and 82.5% tool-selection F1 on Full-Duplex-Bench 3.0, with argument accuracy and end-to-end execution lagging.
Related
Developments·Paper
A frontend-backend architecture for tool calls in full-duplex speech models
An arxiv paper (2026-09-16) proposes a frontend-backend architecture letting full-duplex speech-to-speech models make tool calls. A duplex speech-to-text frontend emits a delegation token and streams ASR transcripts to a text-based backend LLM, whose results are injected back and spoken via streaming TTS. This keeps natural turn-taking and low latency while adding agentic tool-call capability — relevant to anyone building voice agents.
An arxiv paper (2026-09-16) proposes a frontend-backend architecture letting full-duplex speech-to-speech models make tool calls. A duplex speech-to-text frontend emits a delegation token and streams ASR transcripts to a text-based backend LLM, whose results are injected back and spoken via streaming TTS. This keeps natural turn-taking and low latency while adding agentic tool-call capability — relevant to anyone building voice agents.
0 engagement1 sourcearxivWed, Sep 16, 2026, 07:01 PM
Releases·Model Release
Tyger: open-weights streaming ASR + full-duplex turn-control in one model
nineninesix released Tyger on Hugging Face, a small open-weights model that handles full-duplex turn-control for voice agents every 80 ms with no lookahead. It classifies each moment as answer, stop, or chime in, hearing both mic input and the agent's own playback. The accompanying tyger-s2s stack is a complete local voice agent with tool calls and ~0.7 s response latency on an RTX 5090.
nineninesix released Tyger on Hugging Face, a small open-weights model that handles full-duplex turn-control for voice agents every 80 ms with no lookahead. It classifies each moment as answer, stop, or chime in, hearing both mic input and the agent's own playback. The accompanying tyger-s2s stack is a complete local voice agent with tool calls and ~0.7 s response latency on an RTX 5090.
41 engagement1 sourceredditFri, Oct 16, 2026, 05:42 PM
Developments·Paper
VoiceChat-TTS: A Low-Latency Continuous Speech Synthesis Model for Interactive Agents
Researchers propose VoiceChat-TTS, a low-latency, continuous, and streamable text-to-speech model for interactive agents. It is driven directly by LLM text tokens, enabling real-time adaptability like user barge-in. The model addresses the trade-off between latency and speech quality in duplex speech-to-speech systems.
Researchers propose VoiceChat-TTS, a low-latency, continuous, and streamable text-to-speech model for interactive agents. It is driven directly by LLM text tokens, enabling real-time adaptability like user barge-in. The model addresses the trade-off between latency and speech quality in duplex speech-to-speech systems.
0 engagement1 sourcearxivThu, Aug 13, 2026, 11:37 PM
Developments·Benchmark
APEX-Voice: Can Voice Agents Complete Professional Workflows Through Full-Duplex Interaction
Researchers introduce APEX-Voice, a benchmark of 120 interactive professional workflows evaluating whether full-duplex voice agents can complete delegated work, not just converse fluently. Five frontier real-time voice agents were tested; none exceeded 25% Pass@1 and the best Reliable@3 was 10.8%. The results show fluent speech does not translate into dependable professional task completion, with stateful coordination the dominant failure point.
Researchers introduce APEX-Voice, a benchmark of 120 interactive professional workflows evaluating whether full-duplex voice agents can complete delegated work, not just converse fluently. Five frontier real-time voice agents were tested; none exceeded 25% Pass@1 and the best Reliable@3 was 10.8%. The results show fluent speech does not translate into dependable professional task completion, with stateful coordination the dominant failure point.
0 engagement1 sourcearxivMon, Sep 28, 2026, 11:53 AM
Releases·Model Release
NVIDIA releases NemotronLabs VoiceChat-11B full-duplex model on Hugging Face
NVIDIA has released the NemotronLabs VoiceChat-11B model on Hugging Face, featuring full-duplex voice interaction. The model enables real-time, bidirectional conversation, allowing users to speak and listen simultaneously. This release advances voice AI capabilities for developers building conversational interfaces.
NVIDIA has released the NemotronLabs VoiceChat-11B model on Hugging Face, featuring full-duplex voice interaction. The model enables real-time, bidirectional conversation, allowing users to speak and listen simultaneously. This release advances voice AI capabilities for developers building conversational interfaces.
249 engagement1 sourceredditMon, Aug 3, 2026, 10:24 PM