inclusionAI releases MingTok-Audio: the first unified continuous speech tokenizer
Original title:inclusionAI/MingTok-Audio
AI summary
The inclusionAI team has launched MingTok-Audio, the first unified continuous speech tokenizer that effectively integrates semantic and acoustic features, suitable for speech understanding and generation tasks. The model is based on a purely causal Transformer architecture, removes convolutional layers to improve efficiency, and uses VAE for continuous feature modeling to achieve high-quality audio reconstruction. In speech reconstruction performance, its frame rate is 50, with PESQ reaching 4.21 and 4.04 on the SEED-ZH and SEED-EN test sets, SIM of 0.96, and STOI of 0.98, significantly outperforming comparison models. In downstream ASR tasks, it achieves lower error rates on multiple dialect datasets; for example, WER on the Hunan Minnan dataset is as low as 9.80%.
Key Takeaways
- The inclusionAI team has launched MingTok-Audio,the first unified continuous speech tokenizer that…
- The model is based on a purely causal Transformer architecture,removes convolutional layers to…
- In speech reconstruction performance,its frame rate is 50,with PESQ reaching 4.
- 21 and 4.
- 04 on the SEED-ZH and SEED-EN test sets,SIM of 0.
Quick Facts
- Published
- Updated
- Source
- 蚂蚁 inclusionAI:GitHub 新仓库
- Category
- Model Releases
- AI score
- 57
蚂蚁把语音 tokenizer 做到了 PESQ 4.2 的离谱分数,比第二名翻了快一倍,做语音理解和生成的团队值得拿这个当新 baseline 跑一下。
Source & copyright
This article is based on curated titles and AI summaries from the AIHOT public API, with attribution and links back. Original copyright remains with the original publishers and authors; this site does not mirror full texts.
How to Cite
When citing this article, keep the title, date, and permalink, and credit Orange News as the source.
Orange News: inclusionAI releases MingTok-Audio: the first unified continuous speech tokenizer. September 29, 2025. https://new.nat6.net/en/news/cmorb7ik40079slhfn285yff9/