hotwirednews Get GrokBot

Week 7

9–15 Feb 20262026-W07 · 23 filings · 18 companies · 11 net-new

The Speak layer led the week with 5 of 23 filings, from NineNineSix, MOSI, Ant Group, Daily / Pipecat and 1 more.

Hearspeech-to-text4

DeepgramModel providermodel release2026-02-12

Deepgram adds Hebrew, Persian, and Urdu to Nova-3

Nova-3 gains three right-to-left production languages in one API path, reducing the need to stitch separate vendors for those markets; Deepgram provides no language-specific accuracy figures in the post.

Details2 sources · 1 X post

Deepgram made Hebrew, Persian, and Urdu production speech-to-text models available on Nova-3.

Nova-3 Hebrew, Persian, and Urdu

🚀 Hebrew, Persian, and Urdu are now available on Nova-3, extending coverage across the Middle East and South Asia. Production-grade monolingual speech-to-text with streaming, batch, and Keyterm Prompting – built for global voice applications. Link to Blog: https://t.co/AjTPPlYhp1

https://x.com/DeepgramAI/status/2021994579403362678
VCONICPlatformpartnership2026-02-10net-new

VCONIC partners with Speechmatics for healthcare and FS conversation intelligence

Second Speechmatics enterprise wiring in the same week (before Boost.ai on 12 February) — conversation intelligence in healthcare/FS rather than a full CCaaS agent stack.

Also Use

Details1 source

VCONIC and Speechmatics announced a strategic partnership to bring Speechmatics speech recognition into VCONIC's conversation-intelligence platform for healthcare and financial-services use cases.

VCONIC × Speechmatics · vendor Speechmatics

SpeechmaticsModel providermodel release2026-02-10

Speechmatics launches multilingual Medical Models for clinical speech-to-text

Clinical terminology is the differentiator: Speechmatics reports 96% medical keyword recall and 50% fewer medical-term errors, while the blog’s comparison remains a vendor claim.

Details2 sources · 1 X post

Speechmatics launched Medical Models for clinical speech-to-text, reporting 93% general accuracy and availability in Spanish, French, Dutch, and Finnish alongside English.

Speechmatics Medical Models

Generic speech models fail on terms like "hyponatremia." 👎 Our Medical Models hits 93% accuracy on clinical terminology, and it's now available in multiple languages 🌎. How we did it: https://t.co/WVjncN4972 ... and how it fared against CRO @RiquiHerreros 🥊 See it at Booth 2207 @theviveevent #ViVE2026

https://x.com/Speechmatics/status/2021123336332251472
Daily / PipecatPlatformlaunch2026-02-10

Pipecat 0.0.102 adds OpenAI Realtime STT and Sarvam saaras:v3

Realtime transcription without a full S2S session becomes a first-class STT service.

Details3 sources

pipecat-ai 0.0.102 adds OpenAIRealtimeSTTService, Sarvam saaras:v3 STT modes, and STTMetadataFrame for P99 latency broadcast at pipeline start.

pipecat-ai 0.0.102

Speaktext-to-speech5

NineNineSixModel providermodel release2026-02-14net-new

NineNineSix open-sources KaniTTS-2, a 400M TTS model with voice cloning and its pretraining code

Predecessor to Gepard 1.0 (2026-07-07-nineninesix-gepard-1-0, W28), which moved to an Apache 2.0 Qwen3.5 backbone; the LFM 1.0 licence on these cards conflicts with the launch post, so check terms before shipping it.

Details4 sources · 1 X post

English and Portuguese checkpoints of a 400M-parameter text-to-speech model that runs in about 3 GB of VRAM, released with the full pretraining code so teams can train a voice for their own language.

KaniTTS-2

We just open-sourced KaniTTS2 — a 400M param text-to-speech model that runs in 3GB VRAM with voice cloning support.

https://x.com/defoemark/status/2022752293041639452
MOSIModel providermodel release2026-02-11net-new

MOSS-TTS Family opens five speech and audio generation tools

The two-character streaming trigger makes MOSS-TTS-Realtime the family’s voice-agent option; MOSI’s performance superlatives are its own claims.

Details2 sources · 1 X post

MOSI offers voice cloning, dialogue generation, one-shot timbre creation, streaming TTS, and sound effects through its open family.

MOSS-TTS Family

The era of “childhood” for TTS is over. Introducing the MOSS-TTS Family — flagship cloning, dialogue, one-shot timbre, real-time streaming, and sound effects.🙌🙌🙌 The lineup: MOSS-TTS Flagship: Industry-leading zero-shot voice cloning. Precise control over duration and Pinyin, capable of generating 1 hour of speech. MOSS-TTSD-v1.0: A new standard for dialogue generation. Fully optimized for conversational scenes and low-resource languages. Best-in-class performance across all evaluations. MOSS-VoiceGenerator: One-shot timbre generation. Create voices from a single sentence with complex instruction handling. MOSS-TTS-Realtime: Built for the next era of Voice Agents. Synthesis starts after just 2 characters for instant response. MOSS-SoundEffect: Text-to-audio sound effects to expand your creative toolkit. 🔥 Try it now: https://studio.mosi.cn/voice-synthesis 💻 Deploy (GitHub): https://github.com/OpenMOSS/MOSS-TTS 🔌 API Docs: https://studio.mosi.cn/docs/moss-tts #MOSS #AI #TextToSpeech #TTS #OpenClaw #Agent #OpenMOSS #Opensource #VoiceAgent

https://x.com/MosiAI_Official/status/2021524678947340317
Ant GroupModel providermodel release2026-02-11net-new

Ant Group releases Ming-flash-omni 2.0 with its Ming-omni-tts voice core

Ant claims 93% Cantonese dialect-control accuracy and 46.7% emotion-control accuracy against CosyVoice3 (vendor figures); the same Ming line later produced W43's dedicated Ming-3.0-tts (2026-10-19-ant-group-ming-3-0-tts), and MOSI's MOSS-TTS Family (2026-02-11-mosi-moss-tts-family) shipped open TTS weights the same day (analytical).

Also ThinkSee

Details4 sources · 2 X posts

The 100B-parameter open-source omni model generates native speech, audio, and music, while its Ming-omni-tts voice core adds cloning and attribute control.

Ming-flash-omni 2.0 and Ming-omni-tts

One for all, and all for one 🧧 Introducing Ming-flash-omni-2.0: A specialist in every domain, unified as a capable generalist. A gift from Ling =) - Unified Acoustic Synthesis: Speech, audio, and music combined for unbounded creativity; - "Seeing" to "Knowing": Moving beyond input to true deep semantic understanding; - Native Visual Fusion: Seamless generation, editing, and segmentation;

https://x.com/AntLingAGI/status/2021608733449802066

🎙️Ming-omni-tts-16.8B-A3B, Ming-omni-tts-0.5B, are the Voice Core of Ming-flash-omni-2.0, ready for: -Short-video creator looking for the perfect voiceover -Podcaster turns a monologe into a dynamic duo -Dev looking to integrate a voice assistant into your OpenClaw project

https://x.com/AntLingAGI/status/2023776486982115734
Daily / PipecatPlatformlaunch2026-02-10

Pipecat 0.0.102 adds Resemble TTS and cross-service TTS context IDs

Context IDs let word timestamps and interruptions stay aligned across streaming TTS providers.

Details3 sources

pipecat-ai 0.0.102 adds ResembleAITTSService and a context_id TTS tracking system across 30+ TTS implementations, plus Sarvam bulbul:v3-beta and OpenAI marin/cedar voices.

pipecat-ai 0.0.102

ElevenLabsPlatformlaunch2026-02-10

ElevenLabs adds Expressive Mode to ElevenAgents

Expressive Mode pushes ElevenAgents’ value beyond transcription and turn handling into delivery style; the evidence is a company demonstration rather than an independent conversational-quality test.

Also ThinkUse

Details2 sources · 1 X post

ElevenLabs released Expressive Mode for ElevenAgents to make agent speech more expressive in customer conversations.

Expressive Mode for ElevenAgents

Introducing Expressive Mode for ElevenAgents - voice agents so expressive, they blur the line between AI and human conversations. This is an unedited recording of an agent empathizing with a customer at peak frustration. https://t.co/QT6abvmbir

https://x.com/ElevenLabs/status/2021237336793657447

Showavatars, video2

Daily / PipecatPlatformAPI2026-02-10

Pipecat 0.0.102 adds HeyGen LiveAvatar sandbox and video settings

Developers can exercise LiveAvatar without production session billing defaults.

Details3 sources

pipecat-ai 0.0.102 adds is_sandbox and video_settings on LiveAvatarNewSessionRequest, and defaults LiveAvatar session mode to LITE with VP8.

pipecat-ai 0.0.102

LemonSliceModel providerpartnership2026-02-10net-new

LemonSlice plugs real-time avatars into LiveKit Agents

Company-side primary for the avatar plugin pattern this corpus already tracks from the LiveKit Agents and Pipecat changelogs.

Also Orchestrate

Details1 source

LemonSlice’s LiveKit integration puts its real-time avatars into a LiveKit Agents pipeline so developers keep their own VAD, STT, LLM, and TTS choices.

LemonSlice for LiveKit · vendor LiveKit

ThinkLLMs, speech-to-speech4

KyutaiModel providermodel release2026-02-12

Kyutai open-sources Hibiki-Zero for real-time multilingual speech translation

Voice-preserving translation across four European source languages makes Hibiki-Zero a direct open alternative to a cascaded translation stack; the launch materials do not give a production SLA.

Also HearSpeak

Details4 sources · 1 X post

Kyutai released Hibiki-Zero, an open-source real-time speech-translation model that translates French, Spanish, Portuguese, and German into English while transferring the speaker’s voice.

Hibiki-Zero

We're releasing Hibiki-Zero, a new real-time and multilingual speech translation model that can translate 🇫🇷French, 🇪🇸Spanish, 🇵🇹Portuguese and 🇩🇪German to English: accurate, low-latency, high audio quality, with voice transfer. And best of all: open-source. https://t.co/aAfFu4i2lY

https://x.com/kyutai_labs/status/2022007408898511113
Daily / PipecatPlatformlaunch2026-02-10

Pipecat 0.0.102 adds automatic LLM context summarisation

Long sessions can stay within context windows without custom summariser processors.

Details3 sources

pipecat-ai 0.0.102 adds context summarisation that compresses conversation history at token or message limits, with summary request/result frames.

pipecat-ai 0.0.102

UltravoxModel providermodel release2026-02-09net-new

Ultravox highlights v0.7 on a practical voice-agent evaluation

The evaluation targets the workflow behaviours that component speech benchmarks omit, but the reported scores are not an independent production measurement.

Also HearSpeak

Details2 sources · 1 X post

Ultravox v0.7 is presented as a speech-native model for multi-turn voice-agent tasks including tool calling, knowledge-base grounding, and instruction following.

Ultravox v0.7

Ultravox v0.7 outperforms *every other major model* on both intelligence and latency for voice AI in the new AIEWF benchmark from @trydaily: - Ultravox v0.7: 97.7% - GPT-4.1: 94.9% - Gemini-2.5-Flash: 93.7% - GPT-4o: 91.2% - GPT-Realtime: 86.7% - Gemini-Live: 86% The benchmark tests tool calling, knowledge base grounding, and instruction following in a series of multi-turn LLM conversations. https://t.co/oKx6KwdqM2

https://x.com/ultravox_dot_ai/status/2020943505234043084
KotobaModel providerlaunch2026-02-09

Kotoba adds English–Spanish speech-to-speech translation to its app

The English–Spanish pair extends Kotoba’s realtime interpretation path after its 17 April 2026 Japanese–Korean S2S release; the launch post gives no latency number beyond “super low-latency”.

Also HearSpeak

Details2 sources · 1 X post

Kotoba added low-latency English-to-Spanish and Spanish-to-English speech-to-speech translation to its app.

Kotoba app English–Spanish speech-to-speech translation

Super low-latency speech-to-speed transition🔥English <> Spanish is now available on Kotoba app!! Try it out!

https://x.com/kotoba_tech/status/2020997461310124479

Orchestrateframeworks, evals3

Daily / PipecatPlatformlaunch2026-02-13

Daily releases open STT benchmark measuring TTFS and semantic WER

Gives operators a voice-agent-shaped STT scorecard — latency tails plus LLM-facing accuracy — instead of lab WER alone, and wires the same TTFS numbers into Pipecat's context aggregator.

Details2 sources

Daily and Pipecat open-sourced an STT benchmark that scores realtime voice-agent transcription on Time To Final Segment latency and Semantic WER across 10 services on 1,000 real-speech samples.

STT benchmark for voice agents

TavusModel providermodel release2026-02-10

Tavus releases Raven-1 audio and video perception model

Raven-1 adds audio-and-video perception to Tavus’ conversational stack; the launch is a model announcement, not evidence that every Tavus avatar product now exposes the capability.

Also HearSeeThink

Details3 sources · 1 X post

Tavus released Raven-1, an audio and video perception model for understanding emotion, intent, and context in conversational AI.

Raven-1

Most conversational AI understands words, not people. Introducing Raven-1, our audio and video perception model that gives AI the ability to understand emotion, intent, and context the way humans do. https://t.co/yUed7cKjWL

https://x.com/tavus/status/2021260852083708067
LokutorPlatformlaunch2026-02-10net-new

Lokutor open-sources its Go voice-agent orchestrator with full-duplex audio and barge-in

Teams writing Go get an open duplex voice loop outside the Python frameworks (Pipecat first, then LiveKit Agents), though its turn handling is energy-based VAD rather than a learned turn model; Lokutor later ships a Pipecat TTS plug-in (2026-06-01-lokutor-pipecat-tts).

Details2 sources

Lokutor released the Go orchestrator behind its own voice agents as open source, handling simultaneous capture and playback, RMS-based voice activity detection, and user interruptions across pluggable STT, LLM, and TTS providers.

Lokutor Orchestrator for Go

Connecttelephony, channels2

Boost.aiPlatformpartnership2026-02-12net-new

Boost.ai partners with Speechmatics for regulated European Voice AI

Nordic conversational-AI leader wires production ASR from Speechmatics for regulated EU/US voice — dialect and on-prem posture matter more than a new foundation model.

Also HearUse

Details1 source

Boost.ai and Speechmatics announced a partnership integrating Speechmatics speech recognition into Boost.ai's conversational AI platform for banking, insurance, healthcare, and public-sector voice deployments in Europe and the US.

Boost.ai × Speechmatics Voice AI · vendor Speechmatics

GenesysPlatformlaunch2026-02-10net-new

Genesys unveils Agentic Virtual Agent powered by large action models

CCaaS-native agentic resolution with a LAM partner lands in February — months before the September APT-2 and ElevenLabs/Deepgram voice pack already on file.

Also Orchestrate

Details1 source

Genesys announced Genesys Cloud Agentic Virtual Agent, pairing Scaled Cognition's APT-1 large action model with Cloud governance so virtual agents can plan and execute multistep CX workflows across enterprise systems.

Genesys Cloud Agentic Virtual Agent (LAM) · vendor Scaled Cognition

Useagents in market3

ElevenLabsPlatformlaunch2026-02-11

ElevenLabs launches a government initiative for multilingual voice and chat agents

A government-specific offer turns the general agent stack into a vertical procurement and deployment proposition, with voice access as the front door to public services.

Also SpeakOrchestrate

Details2 sources · 1 X post

ElevenLabs launched a public-sector offering for always-on, multilingual voice and chat agents serving government information, benefits, and crisis-response workflows.

ElevenLabs for Government

Introducing ElevenLabs for Government - a dedicated initiative to help public sector organizations deliver always-on services, expand equitable access, and strengthen crisis response.

https://x.com/ElevenLabs/status/2023457485403623646
AmulEnterpriselaunch2026-02-11net-new

Amul launches Sarlaben, a Gujarati voice assistant that dairy farmers reach by phone call

Feature-phone farmers get the same herd-specific advice as app users because the channel is an ordinary call, with Plivo and BHASHINI named on the voice path; the 20-language expansion is undated, and BHASHINI's open VoicERA stack (2026-02-18-bhashini-voicera, W08) launched a week later with the same EkStep partner, without saying Sarlaben runs on it.

Also Connect

Details6 sources · 2 X posts

Amul AI's Sarlaben answers the cooperative's 3.6 million milk producers on cattle health, vaccination, feeding, breeding, and subsidies, through a phone line for feature-phone users as well as the Amul Farmer app.

Amul AI (Sarlaben) · vendor Plivo and BHASHINI · delivered by EkStep Foundation

With the integration of the BHASHINI solution ... Amul AI enables seamless multilingual and voice enabled access

https://x.com/_BHASHINI/status/2021501207613931689

enables access of information through phone and Amul farmer app to 3.6 million dairy farmers in Gujarati through the digital assistant Sarlaben

https://x.com/Amul_Coop/status/2023653646140141779
Simple AIPlatformfunding2026-02-10net-new

Simple AI raises a $14M seed led by First Harmonic for voice agents that sell to consumers

The round paid for the in-house stack that later shipped as Tango (2026-09-29-simple-ai-tango, W40); the 30% upsell figure is a vendor claim with no named customer in the release.

Details3 sources

Simple AI, a Y Combinator company whose voice agents take inbound sales and support calls for consumer brands, raised a seed round led by First Harmonic.