Microsoft and ByteDance join the full-duplex race

Voice AI weekly digest

Events

• Agora runs a hands-on Voice AI workshop in San Francisco (Aug 13, San Francisco, Voice AI Space)

Top Updates 💪

• Microsoft tests MAI-Realtime, its first native full-duplex voice model that listens while speaking across 17 languages, cutting Copilot’s reliance on OpenAI. (Startup Fortune)

• ByteDance launches SeedRealtime, a full-duplex model that watches, listens, and speaks at once, now live in the Doubao app’s phone-call feature. (ByteDance)

• Wispr Flow launches Notetaker, a Mac meeting assistant that captures system audio bot-free, cleans transcripts, and auto-creates tasks. (9to5Mac)

• Omilia raises $67M Series B to expand its voice-first agentic CX platform, having grown live ARR more than 10x to over $60M since Series A. (CMSWire)

• Five9 lands a ~$100M contract with a Fortune 100 financial services firm migrating off on-prem, won in partnership with Google Cloud. (Pulse 2.0)

• Salesforce launches Japanese Agentforce Voice using Kotoba’s Koto model for low-latency speech that handles honorifics and context-dependent nuance. (IBTimes)

• Sierra lands CarMax for inbound sales call AI, boosting call resolution and cutting unresolved calls since its May deployment. (CMSWire)

• Yellow.ai goes public via a $550M SPAC merger, aiming to buy legacy BPOs and convert them into AI-native operations. (CMSWire)

• 8x8 posts record revenue as AI adoption surged 121% year over year, with over 2,900 AI agents built on its platform. (IT Brief)

• 3CLogic wins a major hospital system to modernize its IT service desk with voice AI embedded natively in ServiceNow. (PR Newswire)

• Granola faces a class-action lawsuit alleging its bot-free notetaker records meetings without all-party consent and trains AI on them by default. (Computerworld)

• ElevenLabs deepens its Tokyo push, adapting its voice platform for the Japanese market as APAC expansion accelerates. (btrax)

• Orvera AI is shortlisted by Everest Group in its voice AI agents for CXM spotlight, cited for strength in banking, insurance, and healthcare. (MarTech Series)

• Telnyx makes the case for owned voice AI infrastructure, arguing single-backbone control of telephony, compute, and inference beats stitched multi-vendor stacks. (Under30CEO)

---

Last Week

Engineering Corner 😎

• OpenAI details how it built GPT-Live, a full-duplex realtime system with a delegation layer that hands hard tasks to GPT-5.5 while keeping you talking. (OpenAI)

• Google Creative Lab ships an offline Gemma Translator running Gemma 4 E2B on a Raspberry Pi 5, with full source and 3D-print files released. (AI Weekly)

• Cisco IT shares how it modernized voice security with AI, cutting toll fraud 70% and manual investigation effort 60% with behavioral analytics. (Cisco)

• AWS publishes a serverless real-time voice AI pattern for enterprise sales coaching, built on native services for live transcription and analytics. (AWS)

• Microsoft shows live speech-to-text with Foundry Local and C#, streaming raw PCM audio to a local Nemotron model with no API key or cloud. (Microsoft)

• A dev.to tutorial turns a phone call into a formatted email in 80 lines of Python, cleaning up AI voice memos with the Telnyx API. (dev.to)

• A dev.to explainer breaks down modern voice cloning, walking through the speaker encoder, synthesis model, and vocoder pipeline. (dev.to)

• A new paper introduces QazAVSR, a Kazakh audio-visual speech recognition model built on HuBERT and ViT with a 57-hour dataset. (MDPI)

• A study adapts foundation models for Turkic speech-to-speech translation, comparing fine-tuned cascade and direct end-to-end approaches. (MDPI)