Mistral, xAI and Salesforce Ship New Voice Models

Voice AI weekly digest

Top Updates 💪

• Mistral launched Voxtral 3, a full voice agent stack on LiveKit with STT, TTS and sub-second turns, EU-hosted or self-hosted. the-decoder

• xAI released Grok Voice 3.0, a speech-to-speech model at $0.08/min that it says handles Starlink and Tesla calls. xAI

• Salesforce introduced Launa-Omni, a compact voice model that writes a text reply before speaking it. Salesforce

• Sarvam shipped Bulbul v4 Flash, an Indian-language TTS model it claims is 40% faster and sub-250ms. CIOL

• Gnani launched Timbre v3.0, a TTS model with tags for whispers, laughs and shouts across Indian languages. AnalyticsIndiaMag

• SquadStack launched Arth v2, a streaming STT model tuned for 8kHz Indian sales calls at 67ms latency. CIOL

• Otter.ai raised $320M in non-dilutive growth capital from General Catalyst to expand its enterprise business. TechCrunch

• Suki raised $90M to move its ambient scribe into agents for referrals and prior authorizations. TechTimes

• Ex-Plaud exec raised $26M for Ambi, an always-listening AI assistant for iPhone, Mac and Apple Watch. TechCrunch

• SoundHound closed its LivePerson deal, taking its customer base to 25 of the Fortune 100. CMSWire

• Apple acquired assets of Samplab, a Swiss AI audio startup, surfaced via an EU filing. AppleInsider

• Amazon launched Bee Frontier and Eclipse, two always-listening AI wearables starting at $74.99. CNBC

• Phonak launched Roger On Sphere, a remote mic doing real-time deep-neural-network speech separation for hearing devices. Sonova

• Superwhisper and Cohere partnered on a fully on-device meeting recorder to challenge Otter and Granola. BetaKit

• Hedy launched Hedy for Teams, adding on-device transcription and EU data residency for organizations. Yahoo

• Deepgram said it is targeting air-gapped laptops, running its models natively on-device with Dell. SiliconANGLE

• The FBI arrested a former OpenAI employee over a hidden recording device found in an executive conference room. NBC News

• A new report put deepfake fraud at $40B/year, with voice cloning behind about 22% of incidents. SecurityBrief

• Kira AI launched “life operating software” that places calls, books trips and negotiates bills with approvals. OpenPR

• Exotel’s Sachin Bhatia argued voice AI should move to outcome-based pricing, not per-minute billing. NewsBytes

• Vobiz raised $1M to build an AI-native telephony layer targeting sub-80ms carrier latency for voice agents. Inc42

• FonadaLabs pitched reliability-first voice AI with India-hosted models and telephony for regulated sectors. IT Voice

• New research flagged accent bias in voice AI, warning travel firms that transcripts can misrecord commitments. Travel And Tour World

• PolyAI and Otter execs said voice AI lacks its “ChatGPT moment,” citing speed and transcription gaps. TechCrunch

• CEOWORLD argued the next breakthrough is reliability, not how human the voice sounds. CEOWORLD

• CMSWire noted friendly voice bots still fail on pauses, interruptions and broken handoffs. CMSWire

• Boya launched the Notra Nano, a 24g wearable recorder with transcription in 140+ languages and 16-hour battery. Newsshooter

---

Engineering Corner 😎

• Fish Audio released Drama 3, a 35B open-weight MoE TTS model on Hugging Face with word-level direction. MarkTechPost

• Canonical shipped Myna as a tech preview, a local Ubuntu dictation tool supporting Parakeet, Whisper and FunASR. 9to5Linux

• Deepgram ported Nova-3 to AWS Trainium2, processing 11-70% more audio per dollar on batch workloads. AWS

• NYU Stern detailed Viva, a voice AI that runs and grades oral exams at about $0.42 per student. CACM

• A new paper studied TTS and voice-conversion augmentation to improve ASR on banking call-center audio. arXiv

• FinancialIT broke down why bank collections calls are hard, pointing to latency, escalation and compliance. FinancialIT

• NoiseWorks released PloseAssist, a free plugin that removes vocal plosives automatically. Sonicstate