Voice AI moves to Earbuds and Glasses

Voice AI weekly digest

Events 🎤

• Decagon runs Decagon Dialogues 2026 (Oct 1, San Francisco, Decagon)

• Speechmatics and Tuner co-host a voice AI meetup (Oct 1, London, LUMA)

Top Updates 💪

• Meta unveils Ray-Ban Meta Audio, camera-free AI glasses at $349 that pair earbud audio with the Muse assistant, translation, and calls. (Meta)

• Krisp publishes an open Voice Isolation benchmark, testing 11 STT configurations on real audio with a second voice, where word error rate drops 73%. (Hugging Face)

• Google launches Gemini 3.8 Flash TTS and Flash-Lite TTS (Google)

• Qualcomm announces Snapdragon Sound Elite Gen 2, an audio chip that runs on-device AI and Wi-Fi in earbuds and glasses without a phone. (CNET)

• SoundHound introduces OASYS Edge, fully embedded agentic voice AI that runs locally in vehicles and smart devices even offline. (GlobeNewswire)

• Sarvam AI releases Saaras V4, a speech-to-text model covering all 22 Indian languages plus English with sub-150ms streaming. (Analytics Insight)

• Inworld ships Realtime TTS-2, giving developers natural-language control over delivery in live interactions, with a Flash variant at 25ms time-to-first-byte. (Entrepreneur)

• Sela raises $21M for voice AI agents that help originate over $1B in mortgage loans a month, used by six of the ten largest independent mortgage banks. (Yahoo Finance)

• Dextr AI raises $6.7M for a hospitality agent platform whose Daisy voice agent books reservations and takes payments in 90+ languages. (WebWire)

• Krisp brings real-time voice AI to contact centers, installing at the OS level to add translation, accent conversion and voice security on any softphone. (PC Tech Magazine)

• A 60-org coalition backed by the Gates Foundation commits to bringing AI to 3.4 billion people in their own languages within five years. (Gates Foundation)

• A Washington court rules patients can’t access ambient AI recordings, treating AI scribe audio as exempt administrative documentation. (Clinical Advisor)

• Gizmodo argues AI assistants belong in your ears, not on your face, making the case for earbuds over camera glasses on privacy and social grounds. (Gizmodo)

• Telecoms.com argues voice AI inference belongs inside the network, where edge deployment cuts latency and adds native deepfake defense. (Telecoms.com)

• QSR Web says voice AI has rewritten the drive-thru, as chains scale order-taking agents to ease staffing and keep the line moving. (QSR Web)

---

Engineering Corner 😎

• NVIDIA releases Nemotron 3 Diarization, a 100M-parameter open-weight model that tracks up to eight speakers in real time and tops Diarization-Bench. (HackerNoon)

• AWS shows deploying Qwen3-TTS on SageMaker, with 3-second voice cloning and instruction-driven timbre and emotion across 10 languages. (AWS)

• AWS packages WhisperX on SageMaker for word-level, speaker-labeled transcription with alignment, diarization, and subtitle output. (AWS)

• Pariveda built real-time clinical voice notes for Henry Schein One on Amazon Nova and Bedrock, turning chairside talk into structured notes. (AWS)

• A dev.to guide builds sub-400ms voice agents for enterprise telephony, covering VAD tuning, streaming STT, and turn-level latency budgets. (dev.to)

• SayScroll launches a voice-activated AI teleprompter that scrolls to a speaker’s cadence and pauses on hesitation across 60+ languages. (Dispatch)