hotwirednews Get GrokBot

Week 28

6–12 Jul 20262026-W28 · 19 filings · 18 companies · 4 net-new

The Hear and Orchestrate layers led the week, with 5 filings each.

Hearspeech-to-text5

MOSIModel providermodel release2026-07-09

MOSS-Transcribe-Diarize opens long-form speaker-labelled transcription

The 0.9B weights ship under Apache 2.0 with day-zero SGLang, vLLM, and MLX-audio support, six months after MOSI’s January 2026 technical report (arXiv 2601.01554).

Details2 sources · 2 X posts

The Apache 2.0 open 0.9B model produces speaker labels and timestamps in one pass for audio of up to about 90 minutes.

MOSS-Transcribe-Diarize-0.9B

🤗 MOSS-Transcribe-Diarize-0.9B is now open source on @huggingface. Built with an end-to-end audio-to-structured-transcript paradigm: >0.9B open-source ASR model >Apache license 2.0 >128k long-context transcription >Up to ~90-min audio input >Speaker labels + timestamps in one generation >Multi-speaker diarization for meetings, interruptions, and overlapping voices >Hotword biasing for names, terms, and domain-specific vocabulary >~100 token/s on NVIDIA RTX 4090, RTF ~0.017 Thank you @sgl_project @vllm_project @Prince_Canuma @lllucas for day-0 support! 🚀 Github: https://github.com/OpenMOSS/MOSS-Transcribe-Diarize Huggingface: https://huggingface.co/spaces/OpenMOSS-Team/MOSS-transcribe-diarize Technical Report:http://arxiv.org/abs/2601.01554 SGLang-Omni: https://github.com/sgl-project/sglang-omni vLLM: https://github.com/vllm-project/vllm MLX-audio: https://github.com/Blaizzy/mlx-audio

https://x.com/MosiAI_Official/status/2075059157443756245

🤗 MOSS Transcribe series update.@Open_MOSS Alongside MOSS-Transcribe-Diarize-0.9B, we are also updating two models: MOSS-Transcribe-Diarize Pro >Flagship multi-speaker transcription for complex meetings, long audio, multilingual scenarios, and enterprise API use. MOSS-Transcribe >Open-source ASR for challenging English speech, including standard English, diverse accents, quiet speech, and whispers. Built for enterprise meetings, call-center QA, podcast transcription, interview cleanup, and content production.

https://x.com/MosiAI_Official/status/2075126930681270530
CartesiaModel providermodel release2026-07-09net-new

Cartesia ships Ink-2 streaming STT with semantic turn detection

Gives Cartesia an STT SKU beside Sonic TTS, so agent builders can keep turn detection inside one vendor stack instead of pairing Sonic with Deepgram Flux or AssemblyAI.

Also Orchestrate

Details2 sources

Cartesia released Ink-2, its first streaming speech-to-text model for realtime voice agents, with built-in semantic endpointing and English-only coverage at launch.

Ink-2

Alibaba / QwenModel providermodel release2026-07-09

Alibaba upgrades Fun-ASR-Realtime to first-token latency in the low hundreds of milliseconds, beside Fun-ASR-Flash on Alibaba Cloud

Fun-ASR is Tongyi Lab's separate ASR line from Qwen-Audio: Model Studio's release log adds Qwen-Audio-3.0-ASR-Flash on 30 July, three weeks later, so Alibaba sold two ASR families in parallel and pitched this one on Chinese dialect accuracy (Shanghainese 92.4%, Wenzhou 82.7%) rather than English WER. It shipped the same week as Cartesia Ink-2 (2026-07-09-cartesia-ink-2) and MOSI's open MOSS-Transcribe-Diarize (2026-07-09-mosi-moss-transcribe-diarize-open-source).

Details3 sources · 1 X post

Tongyi Lab's real-time speech recognition model now returns its first token in the low hundreds of milliseconds with accuracy approaching the offline Fun-ASR-Flash, which Alibaba says ranks first on Artificial Analysis at 1.7% WER; both are live on Alibaba Cloud Model Studio.

Fun-ASR-Realtime upgrade (with Fun-ASR-Flash)

First-token latency is now down in the low hundreds of milliseconds, and accuracy is getting close to our offline model.

https://x.com/TONGYI_SpeechAI/status/2075094002932896244
WillowModel providermodel release2026-07-07

Willow releases Frontier Pro and Frontier Mini dictation models and makes Mini free and unlimited

Willow made unlimited cloud dictation free on 7 July, about 12 weeks before Fireflies Talk (2026-09-29-fireflies-talk, W40) did the same; the speed and accuracy claims against Deepgram, Wispr Flow, and OpenAI come from Willow's own tests with no published method.

Details3 sources · 2 X posts

Willow replaced its Atlas 1 dictation model with two cloud models that pair ASR with an RL-trained post-ASR edit model: Frontier Pro for paying Pro users and Frontier Mini for unlimited free use.

Willow Frontier Pro and Frontier Mini

Introducing Willow Frontier Mini and Willow Frontier Pro. Frontier Mini gives everyone free, unlimited dictation. Frontier Pro is our most accurate model yet, built for power users and teams who want the fastest and best dictation experience.

https://x.com/WillowVoiceAI/status/2074525110774681830

We just released Willow Frontier Pro, our fastest and most accurate dictation model yet. Faster and more accurate than Deepgram, Wispr Flow, OpenAI, and more. The most seamless voice typing experience is here. We’re live on Product Hunt today.

https://x.com/WillowVoiceAI/status/2074937227067666514
CohereModel providermodel release2026-07-07

Cohere releases Transcribe Arabic, an open speech recogniser for Arabic dialects and Arabic-English code-switching

Cohere's 25.87 average WER, 2.45 points ahead of OmniASR-LLM-7B, is self-reported on one leaderboard; TII's Falcon-ASR (2026-10-06-tii-falcon-asr, W41) later targeted Emirati and Standard Arabic, so the corpus now holds two open Arabic recognisers with no shared benchmark run.

Details3 sources · 1 X post

Cohere Transcribe Arabic retrains the 2B Transcribe architecture on dialect-balanced Arabic, Arabic-accented English, and code-switched speech, with Apache 2.0 weights on Hugging Face and hosted access through the Cohere API or Model Vault.

Cohere Transcribe Arabic (cohere-transcribe-arabic-07-2026)

We’ve built Cohere Transcribe Arabic, the world’s most accurate open-source model for Arabic speech recognition.

https://x.com/cohere/status/2074499759616729149

Speaktext-to-speech4

LiveKitPlatformlaunch2026-07-09

LiveKit Agents 1.6.5 adds expressive mode preview and Rime speed_alpha

Speak-path cluster before the 1.7.0/blog expressive GA narrative.

Details1 source

livekit-agents 1.6.5 lands an early expressive mode feat plus Rime speed_alpha for Arcana/Coda and SmallestAI word timestamps.

livekit-agents 1.6.5

GradiumModel providerfunding2026-07-09

Gradium (Kyutai spinout) reopens seed to ~$100M with Nvidia

About $100M, Nvidia in, still no model drop. Renault is the named customer. Original stealth round was about $70M in Dec 2025.

Details2 sources

Gradium reopened its seed to about $100M total, with Nvidia among new backers, and announced a Bay Area office expansion.

Gradium

NineNineSixModel providermodel release2026-07-07

NineNineSix releases Gepard 1.0, a 555M streaming TTS model under Apache 2.0

A self-hostable TTS priced against hosted vendors: NineNineSix's 24 August comparison lists its API at $5 per million characters against $30 for Cartesia and $50 for ElevenLabs Flash, and the founder's 10 July Pipecat demo runs it in a cascaded agent on a single RTX 4000 Ada.

Details6 sources · 2 X posts

Open-weight text-to-speech that starts speaking as LLM text streams in, with voice cloning from a few seconds of audio, for English, Spanish, Portuguese, and Dutch.

Gepard 1.0

We open-sourced Gepard 1.0 - a 555M streaming TTS that starts talking as text arrives.

https://x.com/defoemark/status/2074568033092161852

Runs on RTX 4000 Ada on @runpod + @pipecat_ai.

https://x.com/defoemark/status/2075453817236132076
SpeechifyModel providerlaunch2026-07-06

Speechify launches Simba 3.2, a streaming-native TTS model

The $6 per-million-character price and leaderboard placements are Speechify claims; Coval’s later 106 ms measurement is a separate observation.

Details2 sources · 2 X posts

Simba 3.2 is an English-only streaming model that Speechify says has lower time to first byte than Simba 3.0.

Simba 3.2

At just $6 per 1M characters, it's also the most cost-efficient model on the leaderboard.

https://x.com/SpeechifyAI/status/2074246369741844644

At just $6 per 1M characters, it's also the most cost-efficient model on the leaderboard.

https://x.com/SpeechifyAI/status/2074276298902679768

ThinkLLMs, speech-to-speech1

OpenAIModel providerlaunch2026-07-08

GPT-Live replaces Advanced Voice Mode in ChatGPT

The API SKU is the 10 Sep filing: GPT-Live-1 at $0.05/min for the voice layer.

Details3 sources · 2 X posts

OpenAI replaced Advanced Voice Mode defaults in ChatGPT with full-duplex GPT-Live models, with interruptions, live translation, and mid-conversation delegation to GPT-5.5-class models.

GPT-Live (ChatGPT)

Introducing GPT-Live, a new generation of voice models for natural human-AI interaction. Rolling out in ChatGPT starting today. You’ll want to turn the sound on for this one.

https://x.com/OpenAI/status/2074907025537224840

GPT-Live is now fully rolled out to all ChatGPT users on Go, Plus, and Pro plans. Free user rollout is in progress. Update to the latest version of the ChatGPT app on iOS or Android to try it out.

https://x.com/OpenAI/status/2075019750569378007

Orchestrateframeworks, evals5

NVIDIAPlatformlaunch2026-07-09

NVIDIA rebuilds Nemotron Voice Agent 2.0 on upstream Pipecat and adds Nemotron 3 Omni examples

The July rebuild makes upstream Pipecat 1.3.0 the base, while 2026-05-28-pipecat-1-3-0-multi-agent-workers records Pipecat 1.3.0's separate release; removing speculative speech processing trades that earlier latency path for the Smart Turn default.

Details1 source

The re-architected blueprint uses upstream Pipecat directly, a unified React client, and recipe-style deployment profiles for multimodal voice-agent examples.

Nemotron Voice Agent v2.0.0

LiveKitPlatformlaunch2026-07-09

LiveKit Agents 1.6.5 adds realtime model fallback adapter and tool call status events

Realtime failover orchestration.

Details1 source

livekit-agents 1.6.5 adds a realtime model fallback adapter and streams tool call status events.

livekit-agents 1.6.5

ElevenLabsPlatformlaunch2026-07-09

ElevenLabs launches ElevenAgents Spotlight for conversation quality monitoring

Plain-language evals plus real-time anomaly detection move post-launch agent quality into the same operating loop as latency and traces, but ElevenLabs gives no independent CSAT or resolution lift for Spotlight.

Details2 sources · 1 X post

ElevenLabs introduced Spotlight, an observation and improvement layer that reviews voice and chat conversations, groups topics, scores quality, and recommends agent changes.

ElevenAgents Spotlight

Introducing ElevenAgents Spotlight - the fastest way to lift resolution and CSAT across every channel. ElevenAgents Spotlight observes every conversation, voice and chat, in realtime - so you always know what's working and how to improve. https://t.co/K25Uvu8bS7

https://x.com/ElevenLabs/status/2075226064129872324
CrestaPlatformlaunch2026-07-09

Cresta Training Simulator: AI customers for human-agent practice

Voice AI here is charged as training infrastructure for humans, which is a different budget line from replacing the IVR.

Details1 source · 1 X post

Cresta launched Training Simulator: AI-simulated customers built from a company’s real conversations, so human agents practice live and are graded with the same QA criteria as production calls.

Training Simulator

Cresta Training Simulator replaces scripted role-play with live, dynamic AI simulations that help human agents practice before going live.

https://x.com/cresta/status/2075206570225414290
Sarvam AIPlatformlaunch2026-07-06

Sarvam introduces an MCP Server for coding agents

A one-configuration MCP path puts Sarvam’s model and tool catalogue inside coding agents, but the 6 July launch does not specify which voice APIs or agent actions the generic server exposes.

Details2 sources · 1 X post

Sarvam introduced an MCP Server that lets coding agents discover and build with Sarvam models and tools through a single configuration.

Sarvam MCP Server

Today, we’re introducing the Sarvam MCP Server. It gives coding agents a standard way to discover and build with Sarvam. All you need to is a add a single configuration, and Sarvam becomes available directly in your coding agent. It already knows the right models and tools, so you can build with them and ship faster. Create a key, copy the config, start building - https://t.co/SPda9a5ZrX

https://x.com/SarvamAI/status/2074116231461011764

Connecttelephony, channels2

TavusPlatformlaunch2026-07-08

Tavus lets PALs join Google Meet calls as participants

Moving a real-time PAL into an existing Google Meet removes a separate deployment surface for meeting agents; the launch describes capabilities and use cases, not customer adoption or outcome data.

Also OrchestrateShow

Details2 sources · 1 X post

Tavus launched a Google Meet integration that lets a PAL join an existing meeting as a real-time participant, notetaker, assistant, or host.

PAL Google Meet integration

Introducing Tavus Google Meet integration. Now a PAL can join a Google Meet, a real participant in the call you were already going to have, taking its turn the way any other guest would. Make and deploy a customized notetaker, personal assistant, or even a meeting host in a few clicks.

https://x.com/tavus/status/2074894605125026072
ComcentPlatformlaunch2026-07-07net-new

Comcent open-sources Community Edition AGPL voice infrastructure for self-hosted SIP trunks

Open-source SIP-side contact-centre plumbing with built-in STT summaries and a voice bot is scarce beside proprietary CCaaS; the later managed Cloud Product Hunt launch (2026-10-09-comcent-cloud-product-hunt) hosts the same stack rather than replacing this self-host path.

Also HearOrchestrate

Details3 sources

Comcent published Community Edition under AGPL-3.0: a self-hosted contact-centre stack that connects a buyer's SIP trunk (Twilio, Telnyx, and peers) and adds browser diallers, queues, recording, transcription, summaries, sentiment, semantic search, and a real-time AI voice bot.

Comcent Community Edition

Useagents in market2

Alpha BankEnterprisepartnership2026-07-10net-new

Alpha Bank puts an ElevenAgents voice agent in its call centre's voice gateway

A second European incumbent put ElevenLabs voice agents at the front of its contact centre within two weeks of NTT Docomo Business's contact-centre tie-up (2026-06-29-ntt-docomo-business-elevenlabs-contact-center, W27), but Alpha Bank gives no volumes or results, and the agent triages to advisers rather than resolving requests.

Details1 source

The Greek bank expanded its ElevenLabs work with a custom-voiced agent that greets callers in Greek and English and routes them to specialist advisers.

ElevenAgents voice agent for the call centre · vendor ElevenLabs

TripadvisorEnterpriselaunch2026-07-09net-new

Tripadvisor’s Vesper AI voice agent wins NiCE Excellence in AI Innovation

Named travel-platform voice agent on Cognigy/CXone with vendor-stated 90% vs 71% sentiment — caveat: production started early 2026; Jul 9 is the award/case wire, not the first live call.

Also ConnectHear

Details3 sources

NiCE recognised Tripadvisor’s Vesper at NiCE World London: an AI voice agent on Cognigy Auto Agent and CXone that handles natural-language intent capture, identity verification, containment, and human handoff.

Vesper AI Voice Agent (NiCE Cognigy) · vendor NiCE