Releases·Tool Release
slotvox: end-to-end spoken language understanding toolkit on GitHub with synthetic-data-only intent/slot pipeline
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.
279 engagement1 sourceSat, Sep 26, 2026, 12:50 PM
AI Overview
- GitHub repo slotvox (279 stars, 2026-09-26) provides an end-to-end SLU toolkit: RVQ neural audio codec, joint intent-and-slot models, streaming partial intents, speech-LLM instruction data pipeline, turn-level eval
- Uses only deterministic synthetic tone/chirp/formant data with by-construction labels; includes no real speech corpora, pretrained models, or benchmark claims
- Core runs on NumPy with a CPU-only optional PyTorch extra (TCN/GRU causal encoder + intent head + frame-level slot head); requires Python >= 3.11 and runs fully offline
- Exports interleaved speech-text instruction JSONL for speech LLMs with near-duplicate detection, quality flags (SNR, duration), and schema-versioned golden tests
- Evaluates intent accuracy, strict/partial slot BIO F1, joint turn accuracy with bootstrap confidence intervals; CLI spans schema/generate/train/predict/stream-demo/eval/report
Related
Developments·Paper
Researchers propose Spoken Function Calling for open-domain SLU
A new arXiv paper introduces Spoken Function Calling (SFC), a semantic understanding perspective that uses structured rule definitions to improve open-domain spoken language understanding in large audio language models. It aims to overcome the limitations of traditional SLU, which struggles with in-context learning for open-domain tasks.
A new arXiv paper introduces Spoken Function Calling (SFC), a semantic understanding perspective that uses structured rule definitions to improve open-domain spoken language understanding in large audio language models. It aims to overcome the limitations of traditional SLU, which struggles with in-context learning for open-domain tasks.
0 engagement1 sourcearxivWed, Aug 5, 2026, 05:50 PM
Developments·Paper
ConversationalVoice: Full-Duplex Speech Data from Real Conversations through Source-Faithful Reconstruction and Conversation-Grounded Expansion
A new arXiv paper presents ConversationalVoice, a pipeline that converts real two-speaker conversation excerpts into three complementary training-data artifacts for full-duplex speech models. It targets the problem that turn-taking, overlap, interruption, and backchannel signals are entangled across speakers in noisy recordings. This matters to practitioners building full-duplex voice models who need interaction-timing-preserving data.
A new arXiv paper presents ConversationalVoice, a pipeline that converts real two-speaker conversation excerpts into three complementary training-data artifacts for full-duplex speech models. It targets the problem that turn-taking, overlap, interruption, and backchannel signals are entangled across speakers in noisy recordings. This matters to practitioners building full-duplex voice models who need interaction-timing-preserving data.
0 engagement1 sourcearxivTue, Sep 8, 2026, 02:28 AM
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
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.
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 sourcearxivFri, Sep 18, 2026, 04:27 PM
Developments·Paper
SpeechSense dataset targets fine-grained speech sentiment analysis
Researchers released SpeechSense, a paralinguistic-focused dataset for fine-grained speech sentiment analysis. It addresses limitations of text-centric pipelines that discard acoustic features like prosody, aiming to improve applications such as recruitment and customer service.
Researchers released SpeechSense, a paralinguistic-focused dataset for fine-grained speech sentiment analysis. It addresses limitations of text-centric pipelines that discard acoustic features like prosody, aiming to improve applications such as recruitment and customer service.
0 engagement1 sourcearxivTue, Aug 18, 2026, 03:51 PM