Voxwright is a new MIT-licensed Python 3.11+ toolbox on GitHub for building and validating application-layer mechanisms that connect end-to-end spoken language models. It provides deterministic replay, injection-safe tool calling, workflow orchestration, and evaluation — all offline, without real model endpoints or GPUs. Practitioners get a reproducible harness for testing voice-agent reliability before wiring up live models.
Researchers introduced AgentCheck, an open-source web workbench that allows developers to reproduce, intervene, and mitigate failures in tool-using LLM agents by injecting 12 types of faults into MCP server responses. It records real tool interactions, replays them with perturbations, and caches matching calls for efficient debugging before deployment.
Researchers introduced AgentCheck, an open-source web workbench that allows developers to reproduce, intervene, and mitigate failures in tool-using LLM agents by injecting 12 types of faults into MCP server responses. It records real tool interactions, replays them with perturbations, and caches matching calls for efficient debugging before deployment.
A developer released VoiceManager, an MIT-licensed low-latency streaming speech-to-text and text-to-speech server running locally over a single WebSocket, alongside a Kotlin demo app, HermesConsole, and a Python client. The system supports turn-based conversations with barge-in, wake word activation, and tool calls, aimed at adding voice to AI agents.
A developer released VoiceManager, an MIT-licensed low-latency streaming speech-to-text and text-to-speech server running locally over a single WebSocket, alongside a Kotlin demo app, HermesConsole, and a Python client. The system supports turn-based conversations with barge-in, wake word activation, and tool calls, aimed at adding voice to AI agents.
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.
Speakrail is an open-source, fully-local voice agent with low latency and full-duplex interaction (interjections supported), claimed to rival GPT-Live on some benchmarks. It combines Voxtral Realtime for speech with a turn-taking head, a microturn-finetuned Gemma 4 12B as the LLM, and Breeze TTS 2, all running on one RTX 4090. It targets users who want voice assistants without cloud dependence, for lower latency and privacy.
Speakrail is an open-source, fully-local voice agent with low latency and full-duplex interaction (interjections supported), claimed to rival GPT-Live on some benchmarks. It combines Voxtral Realtime for speech with a turn-taking head, a microturn-finetuned Gemma 4 12B as the LLM, and Breeze TTS 2, all running on one RTX 4090. It targets users who want voice assistants without cloud dependence, for lower latency and privacy.
slotvox is an open-source, fully offline toolkit for end-to-end spoken language understanding (intent classification + slot filling), released on GitHub. It covers the whole pipeline — synthetic waveform generation, log-mel features, joint intent-slot models, streaming understanding, speech-LLM instruction data export, and evaluation — in one repo. The core runs on NumPy with an optional CPU-only PyTorch extra, and it uses only deterministic synthetic data rather than real speech corpora.
slotvox is an open-source, fully offline toolkit for end-to-end spoken language understanding (intent classification + slot filling), released on GitHub. It covers the whole pipeline — synthetic waveform generation, log-mel features, joint intent-slot models, streaming understanding, speech-LLM instruction data export, and evaluation — in one repo. The core runs on NumPy with an optional CPU-only PyTorch extra, and it uses only deterministic synthetic data rather than real speech corpora.