Voice AI Weekly #14 - 4th Week of October 2026
This week in the global voice AI industry, the release of single-pipeline-based voice models and support for on-premises deployments stood out. Legal disputes surrounding voice data processing in contact center AI were also highlighted. Here is a summary of five news stories closely tied to real-world deployment environments.
1. Mistral Unveils Voxtral 3, Supports Self-Hosted Infrastructure Deployment
On October 20, Mistral announced Voxtral 3, bundling speech recognition and speech synthesis models with a voice agent SDK. The real-time speech recognition latency is under 500 milliseconds, supporting more than 50 languages, speaker diarization, and prompt-based vocabulary correction (Context Biasing). In addition to a global API, it supports deployment on private clouds or on-premises infrastructure, allowing organizations to operate without sending voice data externally.
Source Mistral, Voxtral 3 Announcement / TestingCatalog Report
2. ElevenLabs Launches Speech-to-Speech Model Eleven Realtime
ElevenLabs unveiled its first speech-to-speech model, Eleven Realtime, on October 20. It is divided into two versions: Mini for lightweight conversations and Max for tasks with heavy tool calling, with average Time to First Audio (TTFA) of 0.62 seconds and 0.57 seconds, respectively. It consolidates the previously separated pipeline of STT, turn detection, LLM, and TTS into a single model and is compatible with the OpenAI Realtime API. However, specific details on how it handles non-verbal acoustic cues such as emotions or tone were not disclosed. As processing pipelines merge into one, layers that interpret the speaker's state and nuances from audio signals are expected to serve as key differentiators.
Source ElevenLabs, Eleven Realtime
3. Block Previews Audio-Native Full-Duplex Model Dessa
On October 23, the AI research team at payments company Block unveiled Dessa, a full-duplex voice model. Trained from scratch solely on human conversational audio without text conversion or pretrained models, it can listen to the other party and continue answering simultaneously. Block stated that it runs locally at faster-than-real-time speeds on environments with a MacBook Pro M4 Max or higher. Currently in a preview phase for waitlist members and a select group of researchers, detailed benchmarks and weights have not yet been released. Processing audio signals directly without passing through text appears to be increasingly taking shape in the research community.
Source Block, Dessa Announcement / PYMNTS Report
4. IBM Releases Open-Source Granite Speech 5.0
On October 21, IBM released Granite Speech 5.0, a 2-billion-parameter speech recognition and translation model, under the Apache 2.0 license. It ranked second among open-weight models with an average word error rate (WER) of 5.06% across nine English public datasets on the Open ASR Leaderboard. A lightweight 470-million-parameter model and updates for meetings were also released. While performance improvements in noisy and far-field microphone environments were noted, detailed validation figures were not included.
Source Hugging Face, Granite Speech 5.0 / Voice AI Space Summary
5. Google Contact Center AI Privacy Lawsuit: Home Depot Claims Dismissed
On October 22, the U.S. District Court for the Northern District of California issued a ruling on the Google Contact Center AI (CCAI) privacy lawsuit. The court dismissed the claims related to Home Depot customers with leave to amend, while referring the claims of Verizon, Uber, and Discover customers to individual arbitration. The plaintiffs claim that Google monitored calls and used them for model training even after users were transferred from AI agents to human agents, which Google denies. The deadline to file an amended complaint for the Home Depot matter is November 12. This serves as a case demonstrating that the very architecture of routing customer service audio to an external cloud can become a legal issue for enterprises.
Source OpenClassActions, Google Contact Center AI Lawsuit Report
If you have any questions or are considering adopting Voice AI, please contact us at contact@holamago.com. We will be back next week with news from the first week of November.