Deepgram Buys Daily; New Voice Models from NVIDIA, Rime and Microsoft

Voice AI weekly digest

Events 🎤

• AAIF hosts the Voice Agents Forum (Nov 5, SF, Luma)

• Rasa runs an Agent Factory workshop (Nov 3, NYC, Luma)

• Twilio brings Assemble to London (Nov 3, London, Luma)

• Twilio’s builder meetup hits Manchester (Nov 4, Manchester, Luma)

• SDxSDSU hosts a Voice Agent Hackathon (Nov 8, San Diego, Luma)

Top Updates 💪

• Deepgram is acquiring Daily, maker of the open-source Pipecat framework, to build an open voice-agent platform. The Next Web

• NVIDIA released a new open transcription model, Nemotron ASR Streaming 0.6B, handling 40 locales in real time under a commercial license. TechCrunch

• Rime launched Prism, a fine-tunable speech-to-speech model enterprises can train on their own calls and run on-prem. The AI Insider

• Microsoft updated Azure AI Speech LLM, adding 25 languages including 11 Indic ones and up to 2x faster transcription. Visual Studio Magazine

• Gnani launched Prisma v3.0, an STT model trained on Indian phone calls with four transcription modes. CXOToday

• ElevenLabs debuted a live speech translation API, in alpha across 90+ languages, with Deutsche Telekom an early user. Slator

• Google DeepMind added native audio to Genie 4, generating sound alongside its world simulations with real-time text steering. TechCrunch

• Ollama brought local TTS to Macs, using Gradium’s Phonon models across five languages on Apple Silicon. 9to5Mac

• Otter.ai moved beyond transcription, adding a Work Mode that preps, coaches and runs post-meeting tasks. Fast Company

• Granola added free voice typing, a Mac dictation tool with no word limit to challenge Wispr. TechCrunch

• Gemini Notebook added mobile audio recording, capturing up to an hour per session on Android and iOS. 9to5Google

• Floo AI launched its voice agent platform, automating inbound and outbound business calls in 20+ languages. Yahoo

• Captio AI offers live speaker-labeled captions, a phone app transcribing 60 languages with on-device voice profiles. Disabled World

• Reviewers covered the camera-free Ray-Ban Meta glasses, an audio-and-AI model that costs more while dropping the camera. Engadget

• NTT built a conversational voice AI that it says responds at least 50% faster than rivals in call-center tests. Telecompaper

• The Golden State Warriors named RingCentral their official agentic voice AI platform for fan engagement. CXM

• Edge Sound Research raised $2M to help robots and vehicles make sense of real-world sound. SiliconANGLE

• Check Point launched audio guardrails, screening a voice agent’s audio for hidden prompt attacks before it acts. Check Point

• A new report pegged the voice AI agents market near $5B in 2026, forecast to grow past $40B by 2036. GlobeNewswire

• Becker’s detailed voice AI ROI in healthcare, as three provider groups tied it to recovered appointments and staff time. Becker’s

• CX Today argued voice AI on Teams is harder than it looks, with handoffs and native integration the real challenge. CX Today

• A report framed trust as voice AI’s bigger test, finding disclosure matters more to callers than a natural voice. Portugal Startups

• CSO Online called AI notetakers an “everywhere” risk, pushing boards toward approved-tool lists and retention rules. CSO Online

• TMCnet asked who audits what voice agents say, arguing firms must log transcripts, consent and disclosures. TMCnet

• TechBuzz argued voice infrastructure is the next bottleneck, as telephony and latency, not models, now limit deployment. TechBuzz

---

Engineering Corner 😎

• MIT profiled Paris Smaragdis, whose lab builds machines that separate and interpret sound. MIT News

• PyTorch and Arm optimized streaming ASR on edge, using ExecuTorch and SME2 to cut processing time on a phone. PyTorch

• An open project ran live ASR on a $5 ESP32-P4, fully offline at about a 4% word error rate. Open Source For You

• Ollama put Liquid’s LFM3 audio on Windows, a 2B local STT/TTS model in a ~2GB download. dev.to

• Edge0 released Audio8 Realistic 0.6B, a compact open TTS model with zero-shot cloning across 23 languages. Hugging Face

• Destined AI released Hurston-1, an open accented-English TTS model built for phone-line audio. Hugging Face

• A walkthrough added speaker labels to captions, pairing WhisperX diarization with WebVTT voice spans. dev.to

• A tutorial wired AI voice into Notion, connecting dictation to productivity-tool workflows. dev.to