Another small 260k params TTS, hard to believe into UTMOS 4.0 based on examples but ok

https://github.com/lab-emi/GrainSpeech

https://lab-emi.github.io/GrainSpeech/

https://arxiv.org/abs/2609.18856

An advanced dictation model from Whispr, an interesting part is GPRO over user corrections

https://wisprflow.ai/canto

Voice agents related benchmarks somehow passed our attention, here are few recent ones:

https://github.com/sierra-research/tau2-bench

https://github.com/ServiceNow/eva

tau-Voice: Benchmarking Full-Duplex Voice Agents on Real-World Domains
https://arxiv.org/abs/2603.13686

EVA-Bench: A New End-to-end Framework for Evaluating Voice Agents
https://arxiv.org/abs/2605.13841

Serious issues with Qwen3-TTS stability, long texts and audio prompts break things

https://arxiv.org/abs/2609.16989

https://x.com/RmdW_W/status/2102032894600401231

Taming Long-form Text-to-Speech
Rongxiang Wang, Berkin Durmus, Aysegul Orhon, Eduardo Pacheco, Atila Orhon

Long-form text-to-speech (TTS) enables multi-turn conversations with consistent prosody and higher quality voice cloning from longer reference audio. Recent open-weights autoregressive TTS models such as Qwen3-TTS and VoxCPM2 attain state-of-the-art word error rate (WER) and speaker similarity (SIM) on short-form prompts but significantly deteriorate when used with long-form prompts. We propose Localized Attention-Constrained Inference (LACI), an inference-only method to detect TTS errors in near real-time, roll back to the error onset and regenerate with temporary guardrails, adding negligible computational overhead. Using LACI, we improve worst-of-N WER across 10 RNG seeds for Qwen3-TTS-0.6B from 35.2% to 3.4% on prompts longer than 1500 words, even surpassing its short-form reliability of 5.4\% on prompts with fewer than 500 words. To demonstrate the efficacy of LACI on voice cloning reliability, we propose a sliding-window version of the SIM metric that we call wSIM. wSIM exposes several novel failure patterns that are not captured by SIM. LACI improves worst-of-N wSIM from 0.01 to 0.47 on 120 seconds of reference audio while reducing the rate of catastrophic generations with WER above 30% from 26% to below 1%

https://github.com/SamsungLabs/samsone

Samsone is a family of open Small Audio Language Models (SALMs) for efficient audio understanding. The release includes Samsone-99M, Samsone-134M, and Samsone-356M checkpoints for both on-device and server-side inference.

https://arxiv.org/abs/2609.21666

Samsone: A Family of Open Small Audio Language Models for On-Device Inference
Piotr Masztalski, Michał K. Grzeszczyk, Olaf Sikorski

The success of Large Audio Language Models has driven the development of massive multimodal networks exceeding billions of parameters. However, the demand for privacy-preserving, low-latency processing has shifted focus toward Small Audio Language Models (SALMs) capable of on-device execution. In this paper, we introduce Samsone, a family of SALMs designed for edge computing. Our core model, Samsone-134M, establishes a new state-of-the-art for its size class across multiple benchmarks. We further explore the scaling laws of SALMs by introducing Samsone-99M and Samsone-356M. Despite their compact footprint, the Samsone family delivers performance competitive with models orders of magnitude larger. To foster open research and reproducibility, we train Samsone on publicly available data. We release the training code, model weights, mobile-optimized checkpoints and provide an open-source Android application to demonstrate real-time on-device inference of Samsone.

Parakeet v3 ternary quantization with good accuracy (178mb) 113xRT on CPU

https://huggingface.co/moondream/parakeet-redux

https://huggingface.co/spaces/Krisp-AI/VoiceIsolation-Benchmark

Benchmark for the speech recognition with background speech

Data. 265 recordings in real offices, four call-center floors and cars. Each work and call-center recording follows a fixed structure: primary alone, both speakers, secondary alone, primary alone. Speakers read scripts to give exact ground truth; environments, devices and background speech are real, with no synthetic mixing. Segments are hand-labeled primary / mix / secondary / noise.

Evaluation. 11 STT configurations across 9 engines, streaming and batch, run on raw audio and after voice isolation. WER is computed with jiwer 4.0.0 at corpus level (errors pooled across files, not averaged). Reference and hypothesis go through the same normalization: regex splitting of alphanumeric tokens, then NVIDIA NeMo WFST text normalization, then lowercasing, contraction expansion, punctuation and filler removal. Perceptual quality is scored with DNSMOS-C on primary-speaker segments only.

Results.
Corpus WER: 23.29% → 6.26%
All 11 configurations improve
Engine spread narrows from 17–37% to 4–8%
Clean phone audio regresses slightly: 3.48% → 3.91%

Limitation. On clean narrowband audio there's little to remove, and isolation removes some of the primary signal. We report it rather than filtering it out. Segment-level labels also let you check for deletions specifically: a model that correctly outputs silence and one that drops primary-speaker words can post similar corpus WER.

Dataset: https://huggingface.co/datasets/Krisp-AI/VoiceIsolation-Benchmark-Dataset
Model outputs: https://huggingface.co/spaces/Krisp-AI/VoiceIsolation-Benchmark

Facebook's Muse Transcribe is 1st on AA leaderboard and 25th on Huggingface leaderboard

https://x.com/EricBezzam/status/2102368894929514711

Some recent advanced TTS evaluation

Blog by @altsoph from Inworld
https://altsoph.substack.com/p/wtf-is-voice-steering

J-HARD-TTS-Eval, a benchmark designed to evaluate the robustness of autoregressive Japanese Text-To-Speech (TTS) models (V2 is coming)
https://github.com/Parakeet-Inc/J-HARD-TTS-Eval

EmergentTTS-Eval: Evaluating TTS Models on Complex Prosodic, Expressiveness, and Linguistic Challenges Using Model-as-a-Judge
https://arxiv.org/abs/2505.23009

Yodas3 was shared yesterday. 55TB, 1.1M hours of speech data

https://huggingface.co/datasets/espnet/yodas3

from previous experience not very useful to be honest

Highly efficient audio inference engine

https://github.com/pegainfer-project/pega-omni

custom kernel for Mimi codec, etc, 128 Personaplex streams in realtime on GB300, should be like 30 streams on 4090

Interspeech 2026 starts today

https://www.isca-archive.org/interspeech_2026/

Let us be brave to read all this. Surprisingly, very few agentic papers.

Good Paper from Interspeech and dataset for game developers

https://huggingface.co/datasets/NCSOFT/Designed-Vocalizations-Dataset

https://ncai-official.github.io/speech/publications/designed-vocalizations-dataset/index.html

Advances in AI-based voice conversion have enabled a wide range of media applications, including films, audiobooks, and games. However, most research and public benchmarks still focus on natural human speech, leaving designed vocalizations such as monster growls and robotic voices underexplored, partly due to the lack of publicly available resources. To address this gap, we introduce the Designed Vocalizations Dataset, created by applying professional vocal effects processing to diverse vocal sources to produce paired original and effect-modified audio. We further provide a standardized test set with explicit seen/unseen splits over source types and preset styles to assess generalization under controlled conditions, together with baseline benchmark results for reproducible evaluation.

New update from PocketTTS, different objective improves things

https://kyutai.org/blog/2026-09-28-pocket-tts-drifting/

You can inject directly into KV cache ;)

Symbiotic Architecture for Post-Hoc Audio Extension of Frozen Language Models
https://arxiv.org/abs/2609.30784

Popular thing in modern LLMs, Mostik is doing similar research, also

Cache-to-Cache: Direct Semantic Communication Between Large Language Models
https://arxiv.org/abs/2510.03215

The Conversational AI Reading Group is back for Fall 2026!

Over the past two years, we've hosted more than 50 talks from researchers in academia and industry working on speech, audio, and conversational AI, and we're excited to continue this season.

We're kicking off this week with Sreyan Ghosh who will talk about building open audio-language models:

Towards Fully Open General Audio Intelligence
Thursday, October 8 | 11:00 AM – 12:00 PM ET
Speaker: Sreyan Ghosh - Google DeepMind
Details, Zoom link, and the full schedule of upcoming talks: https://poonehmousavi.github.io/rg.html
Youtube: https://www.youtube.com/@CONVAI_RG

Our friend @RND_RandoM recently released a cool full-duplex voice agent setup which that rivals GPT-Live

https://github.com/speakrail/speakrail

The pipeline consists of:

STT: Voxtral Realtime with an attached turn head (HF), running on our audio.cpp fork.
LLM: Gemma 4 12B QAT with microturn finetuning (HF). It is chosen because it fits the GPU quite well, has vision support (I want to test it soon), and in general, the Gemma models perform well in real-life tasks, general chatting, etc.
TTS: Breeze TTS 2, patched to run at int8 (GitHub fork), although it can be replaced by any streaming TTS.
The harness itself: it is the glue between all the components, and has many latency-saving measures, like speculative LLM+TTS firing (inspired by HF speech-to-speech).

The cool thing is that Danil took inspiration from several "think while talking" papers (e.g. SHANKS): while you are talking, a base Gemma 4 12B int4 writes thinking notes, which are then passed to the talker. It helps with harder tasks that require more reasoning.

Please try it out