This is quite insightful paper. Transformers are faster than CNN

https://arxiv.org/abs/2601.20094

T-Mimi: A Transformer-based Mimi Decoder for Real-Time On-Phone TTS

Haibin Wu, Bach Viet Do, Naveen Suda, Julian Chan, Madhavan C R, Gene-Ping Yang, Yi-Chiao Wu, Naoyuki Kanda, Yossef Adi, Xin Lei, Yue Liu, Florian Metze, Yuzong Liu

Neural audio codecs provide promising acoustic features for speech synthesis, with representative streaming codecs like Mimi providing high-quality acoustic features for real-time Text-to-Speech (TTS) applications. However, Mimi's decoder, which employs a hybrid transformer and convolution architecture, introduces significant latency bottlenecks on edge devices due to the the compute intensive nature of deconvolution layers which are not friendly for mobile-CPUs, such as the most representative framework XNNPACK. This paper introduces T-Mimi, a novel modification of the Mimi codec decoder that replaces its convolutional components with a purely transformer-based decoder, inspired by the TS3-Codec architecture. This change dramatically reduces on-device TTS latency from 42.1ms to just 4.4ms. Furthermore, we conduct quantization aware training and derive a crucial finding: the final two transformer layers and the concluding linear layers of the decoder, which are close to the waveform, are highly sensitive to quantization and must be preserved at full precision to maintain audio quality.


Previous paper from the same author

https://arxiv.org/abs/2411.18803

Following on this, NVIDIA implements transformer encoder instead of conformer

https://github.com/NVIDIA-NeMo/NeMo/pull/15661

Interesting comments by Desh on TML

https://x.com/rdesh26/status/2054246456635150744

for example Game-Time

Game-Time: Evaluating Temporal Dynamics in Spoken Language Models
https://arxiv.org/abs/2509.26388
.

ASR robusness still a large area to improve

https://github.com/xzf-thu/Mega-ASR

If you decompose prosody you can do many nice things

https://arxiv.org/abs/2605.05927

Minimizing Modality Gap from the Input Side: Your Speech LLM Can Be a Prosody-Aware Text LLM
Wenqian Cui, Xiao-Hui Li, Daxin Tan, Qiyong Zheng, Irwin King

Speech large language models (SLMs) are typically built from text large language model (TLM) checkpoints, yet they still suffer from a substantial modality gap. Prior work has mainly attempted to reduce this gap from the output side by making speech generation more text-like, but the gap remains. We argue that the key remaining bottleneck lies on the input side. We propose TextPro-SLM, an SLM that makes spoken input more closely resemble that of a prosody-aware text LLM. TextPro-SLM combines WhisperPro, a unified speech encoder that produces synchronized text tokens and prosody embeddings, with an LLM backbone trained to preserve the semantic capabilities of the original TLM while learning paralinguistic understanding. Experiments show that TextPro-SLM achieves the lowest modality gap among leading SLMs at both 3B and 7B scales, while also delivering strong overall performance on paralinguistic understanding tasks. These gains are achieved with only roughly 1,000 hours of LLM training audio, suggesting that reducing the modality gap from the input side is both effective and data-efficient.

Finally a good competition timeline, not two weeks to implement everything

https://saigonaihub.com/OneVoiceAIChallenge

Dare to build the next generation of realtime translation devices powered by Edge AI
Presented by Saigon AI Hub and Qualcomm

🟠 24 May - 24 June 2026: Registration period
🟠 July 2026: Technical specification submission
🟠 August - September 2026: Prototype submission
🟠 October 2026: Field testing
🟠 November 2026: Grand finale at VNG Campus

While everyone focuses on latency, there are many measurable aspects of ASR that are easy to evaluate and have a significant impact on user experience. Here are some of them:

* Hallucination rates from noisy inputs.
* Recognition of the short inputs.
* Ability to identify non-speech sounds as music and noises.
* Rare words problem

Some recent training results from our system

https://alphacephei.com/nsh/2026/05/24/asr-details.html

https://github.com/harrrshall/natscore

Modern neural TTS (CosyVoice2, F5-TTS, MaskGCT, Llasa, XTTS-v2, etc.) generates speech that crosses the threshold where the dominant failure mode is no longer artifacts; it's subtle unnaturalness: prosody glitches, expressive overshoot, speaker-clone drift, breath placement, code-switching mismatches. The existing automatic naturalness scorers were not trained on this kind of failure surface.

UTMOSv2 (the VoiceMOS 2024 winner) was trained on read-speech MOS labels. It saturates at high quality and is documented to produce negative correlations with human judgment on conversational and expressive speech (arXiv 2603.01467).

WhiSQA was designed for telecom and speech-enhancement quality (NISQA training data). It is intentionally not a synthetic-TTS scorer.

DNSMOS, NISQA-TTS, and the rest of the legacy stack predate modern neural TTS and lack the distribution coverage.

SpeechJudge-GRM (released Nov 2025) is excellent, but it is a 7B-parameter LALM. ~$0.001 per score on Modal A100. Unusable inside a TTS training loop or for large-scale offline evaluation.

The data that fixes the distribution gap, SpeechJudge-Data, was released in November 2025: 99K human-labeled TTS preference pairs across CosyVoice2, F5-TTS, MaskGCT, Llasa, and others, in en/zh + code-switching, with both regular and expressive splits. As of writing, no clean public artifact combines this data with a small, deployable, CPU-runnable scorer.
NatScore fills that gap.

https://github.com/ASLP-lab/Smart-Glass-Challenge

https://aslp-lab.github.io/SmartGlasses/

Driven by the rapid advancement of Large Language Models (LLMs) and Multimodal LLMs, AI-powered smart glasses are emerging as a next-generation platform for human-computer interaction. Equipped with microphone arrays and cameras, smart glasses naturally capture the wearer’s egocentric (first-person) perspective, enabling hands-free multimodal communication throughout daily life.

However, deploying robust speech-centric interaction systems on smart glasses introduces distinct challenges compared with traditional stationary devices such as smart speakers or handheld devices such as smartphones. Smart glasses operate in highly dynamic acoustic environments, including environmental noise, user-generated motion noise, and speech from surrounding people.

To address these challenges, the SmartGlasses Challenge introduces a new benchmark for evaluating Time-Stamped Speaker-Attributed ASR (TSA-ASR) and Spoken Language Understanding (SLU) in real-world egocentric interaction scenarios, including dyadic conversation, and multi-party meetings.

Things go fundamental

https://github.com/xzf-thu/Audio-Interaction/

Large-scale streaming-audio dataset for audio-LLM / audio-agent training. Each row is a stream: a sequence of audio turns sharing one unified schema. ~2.28M unique audio clips are organised into six task subsets.

https://huggingface.co/datasets/zhifeixie/StreamAudio-2M

Everyone looks for translation these days. Good paper covering the task complexity (naturalness, prosody, content). Omni systems still unrealistic

https://arxiv.org/abs/2606.03241

Benchmarking Speech-to-Speech Translation Models
Alkis Koudounas, Hayato Futami, Quentin Jodelet, Osamu Take, Shinji Watanabe, Emiru Tsunoo

Speech-to-speech translation (S2ST) has advanced rapidly, but offline evaluation lacks a unified protocol: studies report non-overlapping metric subsets, preventing direct comparisons. We introduce COMPASS, a unified and reproducible benchmarking framework integrating 46 metrics across eight dimensions, and deploy it on 1,248 model-language configurations from FLEURS and CVSS, spanning cascaded and end-to-end architectures over ten language pairs. Architectures exhibit complementary strengths: best-vs-worst gaps exceed 30\% on naturalness and speaker preservation but remain within a few points on translation quality, so single-metric rankings systematically misrepresent system quality. Correlation filtering reduces 46 metrics to 10 per direction, with three axes requiring different metrics across XEN and ENX (e.g., TER/UTMOS vs. ChrF++/NISQA-MOS); these subsets preserve rankings (Spearman's ) while cutting evaluation time by . Human validation across dubbing, podcasts, and medical domains shows standalone MOS predictors fail to predict listener preference, while top domain-specific metrics correlate with human judgment (). We release COMPASS as a foundation for domain-aware S2ST evaluation.

Deepmind paper

SURF: Separation via Unsupervised Remixing Flow


https://google.github.io/df-conformer/surf/

https://arxiv.org/abs/2606.04921

The goal of single-channel source separation is to reconstruct K sources given their mixture. In supervised settings where vast amounts of clean source data are available, this challenging, illposed problem has been addressed successfully by generative diffusion and flow-based prior models. However, access to such clean source samples is often limited. To bridge this gap, we present Separation via Unsupervised Remixing Flow (SURF), an unsupervised flow matching approach for source separation that learns directly from observed mixtures. This method relies on a novel combination of state-of-the-art supervised flow matching and regression-based selfsupervised techniques. At a high level, starting from a teacher model, we utilize a “remixing” step to bootstrap the learning of a student flow model from the teacher’s estimates. We provide insights into the objectives optimized by this approach and draw a novel connection to the Wake-Sleep algorithm. Empirical evaluations on image and audio benchmarks demonstrate that SURF establishes a new state-of-the-art, significantly outperforming existing unsupervised methods.

Recent SynSIG seminars finally uploaded

https://www.youtube.com/@isca-synsig

for example

https://www.youtube.com/watch?v=M8n9I9eGyTM