Amazon Ships Nova 2.5 Sonic; Sierra Debuts an AI-Caller Detector

Voice AI weekly digest

Events 🎤

• Twilio and Tessl host a London builder meetup on conversational AI (Oct 14, London, Luma)

• pyannoteAI, Gladia, and Modal dig into the production voice stack (Oct 14, Paris, Luma)

• WarsawJS puts voice AI on stage at Meetup #143 (Oct 14, Warsaw, WarsawJS)

Top Updates 💪

• Amazon ships Nova 2.5 Sonic, a speech-to-speech model on Bedrock with stronger reasoning, tool calling, and lower latency for real-time voice agents. (AWS)

• Sierra launches Fleming-1, a model that scores a caller’s audio in real time to flag whether the voice on the line is an AI agent. (Unite.AI)

• Krisp makes its flagship Voice Isolation model 2x faster with no quality loss. (Krisp)

• Automation Anywhere agrees to acquire Boost.ai, extending from employee service into customer-facing conversational and voice AI. (SiliconAngle)

• Apple acqui-hires Huxe, the personalized-audio startup from ex-NotebookLM developers, licensing its tech and hiring the team. (AI Weekly)

• Sarvam launches Kivi for Mac, a voice productivity app powered by Saaras V4 that dictates and edits across apps in 22+ Indian languages. (Moneycontrol)

• Krisp ranks #2 of 500 on Cybernews’ AI Trustworthiness index, scoring 97/100 with perfect marks on security, transparency, and privacy. (Krisp)

• TII releases Falcon-ASR, a 1.6B Arabic and Emirati speech model it says beats far larger systems on its benchmarks. (FourWeekMBA)

• Google opens its SynthID detector to the public, a website that checks whether media carries Google’s AI watermark. (TechCrunch)

• Google releases AI Edge Foresight, a free offline Mac note-taker that transcribes and answers questions locally. (The Verge)

• Google’s Pixel 11 adds Call for Me, letting Gemini place routine business calls on your behalf with a live transcript. (AFR)

• DeepL Voice makes Fast Company’s Next Big Things in Tech 2026, cited for voice-to-voice translation that preserves each speaker. (PR Newswire)

• A Nasscom-Krisp report finds only 13% of Indian enterprises have scaled voice AI, with 31% in limited production. (Deccan Herald)

• eHealth moves its voice AI from pilot to full-time, using Regal’s Alice to prescreen up to half a million Medicare calls per enrollment window. (Customer Experience Dive)

• EliseAI partners with Nextech to co-develop agentic voice AI for ophthalmology, dermatology, and other specialty practices. (Manila Times)

• AudioShake launches The Refinery, turning overlapping real conversations into speaker-separated AI training data. (Unite.AI)

• Rebel Audio raises a $10M Series A for an AI podcast production platform with translation, clipping, and one-click publishing. (Hollywood Reporter)

• Sonilo raises $11M, led by B Capital, for a model that generates music and sound effects synced to video. (Variety)

• AncSonic debuts an on-device audio AI architecture, its SORC design targeting earbuds and headphones on low-memory silicon. (Benzinga)

• Voice AI lawsuits put contact centers on BIPA watch, as new voiceprint suits against Walmart and Lowe’s extend the theory to customer service. (CX Today)

• OpenAI plans to retire its legacy TTS voices, drawing pushback from creators who built around tts-1 and gpt-4o-mini-tts. (Six Colors)

• The Verge argues we can’t quietly redefine “recording”, as always-on AI hardware blurs what counts as captured. (The Verge)

• A brain implant decodes speech and gestures at once, a UCSF study driving a full-body avatar for two paralyzed participants. (Fast Company)

• Boston builds an AI “dictionary” of rat calls, using ultrasonic mics and models to map infestations while redacting human speech. (Northeastern)

---

Engineering Corner 😎

• Google releases EmbeddingGemma 2, a 740M on-device multimodal embedding model spanning text, vision, and audio under Apache 2.0. (Google Developers)

• LiveKit’s Jesse Hall reframes voice latency as a budget, arguing turn-taking and time-to-first-audio matter more than raw model speed. (BigGo)

• Pindrop breaks down detecting Meta’s Muse voice, a deep dive into spotting the newest synthetic speech in real time. (Pindrop)

• A comparison of MAI-Transcribe, Grok, and Qwen STT weighs accuracy, latency, and pricing across the three streaming models. (Tech Insider)

• HackerNoon walks through a low-latency STT pipeline for dictation, covering streaming, endpointing, and the latency trade-offs. (HackerNoon)

• Salesforce publishes six principles of voice design, with Agent Script examples and before-and-after audio for Agentforce builders. (Salesforce)

• VoxLocal is a minimal local voice agent in Rust, running Whisper, an LLM, and TTS fully on-device with no cloud. (Lavx)

• babytalk runs offline STT and TTS on an ESP32, a fine-tuned on-device speech stack for noisy environments and microcontrollers. (GitHub)

• MemoTask is a voice-first to-do built for accessibility, using ElevenLabs Scribe v2 and a local Gemma model to capture tasks by voice. (dev.to)

• Voices’ CTO argues AI voice training data is now a quality problem, shifting from scraped volume to consented, targeted sets. (TechBuzz)

• An analysis of Europe’s voice AI cost problem breaks down what TTS and STT really cost at contact-center scale. (Medium)

• A paper tackles low-resource Vietnamese logistics ASR, using LoRA and dynamic curriculum learning for outbound-call speech. (MDPI)