Voice AI's Consolidation Begins

Voice AI weekly digest

Two important Voice AI events in the coming weeks:

• Twilio Signal - May 6-7 in SF

• Cerebral Valley Voice Summit - May 6 in SF

Top Updates 💪

• xAI launches Grok Voice Think Fast 1.0, ranking #1 on the tau-voice Bench for full-duplex voice agents and already powering Starlink support with a 20% sales conversion rate. (xAI)

• Anker unveils THUS, the first compute-in-memory AI audio chip, claiming 150x more on-device AI power for noise cancellation in its upcoming Soundcore earbuds. (The Verge)

• SoundHound acquires LivePerson for $43M, combining voice agentic AI with LivePerson’s digital messaging platform that handles one billion customer messages per month. (GlobeNewswire)

• Krisp Voice AI SDK won double Webby Awards for Technical Achievement (LinkedIn)

• Speechmatics delivers on-device STT for Adobe Premiere, transcribing an hour of video in 55 seconds offline with accuracy within 5% of cloud. (TV Technology)

• Nothing launches Essential Voice, an AI dictation tool that cleans filler words and formats speech-to-text system-wide in 100+ languages. (TechCrunch)

• Synthflow AI and 8x8 partner to embed no-code voice AI agents directly into the 8x8 Contact Center platform across 30+ languages. (VentureBeat)

• Google Meet AI note-taking now works for in-person meetings, generating transcripts, summaries, and action items from face-to-face conversations via mobile. (Lifehacker)

• Xiaomi releases MiMo v2.5 TTS and open-sources MiMo v2.5 ASR, a full voice pipeline with voice cloning, voice design, and dialect-aware recognition for the agent era. (Gizmochina)

• Volkswagen will ship voice AI in all China-built cars starting H2 2026, using on-device LLMs from Tencent, Alibaba, and Baidu. (CNBC)

• Newo appoints new CEO after $25M Series A to scale partner-led voice AI infrastructure for MSPs, VoIP providers, and software platforms serving SMBs. (GlobeNewswire)

• Ericsson embeds AI calling and fraud detection into IMS, partnering with Hiya for real-time spam blocking as 86% of unknown calls go unanswered. (Ericsson Blog)

---

Engineering Corner 😎

• Streaming TTS models fail over 60% of sentences containing phone numbers, dates, and prices due to 5-20x less context than batch mode. (Technology.org)

• AI neck sensor turns silent speech into voice by reading microscopic throat muscle movements with a CNN+transformer pipeline from POSTECH. (Digital Trends)

• AWS guide to cost-effective multilingual transcription at scale using NVIDIA Parakeet TDT and AWS Batch. (AWS Blog)

• Ghost Pepper: open-source browser extension for real-time voice transcription and LLM-powered responses. (GitHub)

• Mimi Codec deep-dive on its layered audio compression design for neural speech coding. (LetsDDataScience)

• AssemblyAI showcases configurable STT with tunable turn-taking, medical mode for streaming, and real-time speaker labeling. (TipRanks)