OpenAI has released GPT-Live, a third-generation voice system that eliminates the traditional turn detector by using a full-duplex model capable of listening and speaking simultaneously. This architecture allows for more immediate and natural conversations while maintaining low latency through a new system design optimized for real-time performance.
xAI has announced Grok Voice Think Fast 2.0, a next-generation speech-to-speech voice model designed to enhance intelligence, transcription accuracy, and conversational capabilities. The update builds on its predecessor by introducing meaningful gains in speech reasoning, conversational ability, and tool use reliability.
Google has launched its newest music generation model, Lyria 3.5, within Google Flow Music to help users create richer tracks with greater creative control.
OpenAI has launched GPT-Live-1 and GPT-Live-1 mini to power a rebuilt ChatGPT Voice, introducing full-duplex audio processing that allows simultaneous listening and speaking. The system delegates complex queries to background reasoning models like GPT-5.5, enabling continuous conversation while deeper thinking occurs.
The llama.cpp project released version b10270, introducing support for the Qwen3-TTS model. This update includes a breaking change to the llama-tts binary and implements significant architectural modifications to handle the new text-to-speech workflow.
NVIDIA has made the NemotronLabs-VoiceChat-11B voice chat model available on Hugging Face. The repository is described as supporting full duplex communication capabilities.
A researcher has published the OpenGauntlet research catalog, a public resource containing information on 168 text-to-speech systems and 106 speech-to-text systems. The directories cover open and hosted models, licensing, hardware requirements, languages, capabilities, lifecycle status, source information, and published benchmark data where available.
PolyAI has introduced Dialog-RSN-1, a dialog model that processes raw caller audio directly rather than relying on transcripts. It fuses turn-taking, speech recognition, function calling, and response generation into a single audio-aware system.
Alibaba’s Tongyi Lab has released Qwen-Audio-3.0-TTS, a hosted text-to-speech system available in two variants: Flash for real-time interaction and Plus for high-quality generation. Both tiers are delivered via Alibaba Cloud Model Studio rather than as downloadable weights.
NymphTech is seeking recommendations for high-end voice cloning and speech synthesis models to enhance its AI radio station network, specifically targeting multiple languages and long-form broadcasting. The company has already launched a platform featuring persistent AI personalities like Trinity and Mark, with Trinity having broadcast continuously since June 9.
Hume has introduced Real World VoiceEQ, a benchmark designed to evaluate the human quality of voice interaction by assessing how well systems recognize and produce acoustic information beyond transcripts. The benchmark evaluates over 40 proprietary and open-source models across 15+ dimensions using more than 60 metrics derived from 785,000 TTS and 48,000 STS human ratings.
A developer on the Hugging Face forums seeks to understand how production platforms like Bland.ai, Retell, and Vapi manage prompt orchestration without relying on massive, hand-written system prompts. The author describes their current implementation using FastAPI, Sarvam STT/TTS, and Pipecat SmartTurn V3, noting that manual state-wise injection leads to growing complexity and edge-case failures.
Researchers propose a framework that uses audio-aware large language models (ALLMs) as fine-grained judges to verify target event presence and temporal relations in generated audio. This approach addresses the gap where existing text-to-audio models often fail to follow instructions involving multiple sound events and their order.
A study audits protocol-level shortcuts in large audio-language models (LALMs) used as automatic judges for speech evaluation, revealing that high agreement with human ratings does not guarantee verdicts are grounded in the actual audio. The research examines three deployment protocols: feature-blueprint judging, reference-conditioned judging, and pairwise A/B comparison across six judges and four attributes.
The study investigates using Dynamic Time Warping (DTW) on self-supervised WavLM representations to assess phonetic accuracy, rhythm, and intonation in English and Japanese L2 speech without requiring labeled training data. The basic DTW approach comparing learner speech to native templates exceeds human agreement on holistic and sentence-level phonetic scoring.
The open-source framework Audient provides a local audio perception layer for LLM agents that continuously processes sound and builds a memory of concepts through use. It uses CLAP embeddings, Whisper, Silero VAD, and sqlite-vec to gate, fingerprint, and recognize audio events without requiring constant LLM calls.
Researchers demonstrate that a discrete diffusion language model can transcribe speech by refining a whole transcript in parallel over a small number of denoising steps, rather than using autoregressive decoders.
The audio.cpp project has released version 0.3, introducing five new text-to-speech models: Supertonic 3, MOSS-TTS-Local, MOSS-TTS-Nano, IndexTTS2, and Irodori-TTS. The update also adds GGUF support, which will be rolled out model by model.
Researchers propose Unified Gradient Projection (UGP), a method to mitigate catastrophic forgetting when fine-tuning large pretrained ASR models like Whisper on low-resource languages. UGP constrains parameter updates using reference gradients from language-balanced replay within a unified projection space, effectively equalizing per-language contributions and reducing dominant-language bias.
The author introduces Cicero, an open-source project that provides self-hosted, bidirectional voice interaction for AI agents. Unlike one-way dictation tools, Cicero allows agents to speak back and call users via Telegram, keeping all audio local.