Voice AI Weekly #13 - 3rd Week of October 2026
We have selected five global voice AI news stories for this week. The competition in model speed continued, accompanied by updates on domain adaptation and on-device capabilities.
1. xAI Unveils Grok Voice 3.0
On October 14, xAI released its speech-to-speech model, Grok Voice 3.0. It features a latency of under 1 second and is priced at $0.08 per minute, the same as 2.0. It took first place on Artificial Analysis's Speech-to-Speech Index with a score of 86.9, compared to 81.3 for 2.0. The Starlink call center handles about 20,000 cases a day, with approximately 70% reportedly resolved without human intervention.
Source: xAI, Introducing Grok Voice 3.0
2. Speechmatics Launches Oak, a Healthcare-Specialized Speech Recognition Model
On October 14, Speechmatics unveiled Oak, a speech recognition model for healthcare. It supports 55 languages and can understand language switches within a single sentence. In internal evaluations, it transcribed medical terminology with approximately 91% accuracy, compared to an average of 74.3% across eight competitors including AWS, Google, Microsoft, and Deepgram. All comparisons were conducted under default settings and reflect internal company research findings.
Source: GlobeNewswire, Voice AI model built to accurately capture real multilingual medical conversations
3. NatWest Tunes NVIDIA Nemotron ASR for British Accents
NatWest has tailored NVIDIA's open-source Nemotron ASR for British accents and financial terminology. Around 90 employees read about 3,000 banking questions and also recorded them rephrased in their own words. Anonymized contact center transcripts were subsequently added. The data was kept within NatWest's environment, and customer data was not shared with NVIDIA. As a result, misrecognized words were reduced by 20%. This voice capability is utilized in the spending analysis service, Spending Insights, and is available to 75% of RBS customers.
Source: The Paypers, NatWest and NVIDIA cut speech AI errors by 20 percent for UK banking
4. Tether Unveils On-Device Speech Recognition TranscribePsy
On October 14, Tether unveiled QVAC TranscribePsy, a speech recognition model with 254 million parameters. It supports 25 European languages on laptops and 8 languages in smartphone configurations. According to Tether's internal evaluation, its Open ASR Leaderboard WER was 8.02%, and it processed 10 minutes of audio in 1.65 seconds on an iPhone 17 Pro Max. It is licensed under Apache 2.0 and has been released for research and educational purposes. These figures have not been independently verified.
Source: crypto.news, Tether launches offline AI speech model for smart glasses and robots
5. Ant Group Open-Sources Ming-3.0 Speech Models
On October 20, Ant Group open-sourced Ming-3.0, a suite of five speech models. Covering speech synthesis, audio understanding and dialogue, and speech recognition, they can be used individually or combined. The audio understanding model summarizes and analyzes long recordings while retaining speaker, emotion, and background sound information. The license is not specified in the article, requiring verification from the original source.
Source: TechNode, Ant Group open-sources Ming-3.0 speech models with voice design and audio understanding
If you have any questions about voice AI implementation or sentiment analysis, feel free to contact us at contact@holamago.com. We will return next week with another roundup of new voice AI news.