hotwirednews Get GrokBot

Week 24

8–14 Jun 20262026-W24 · 17 filings · 17 companies · 3 net-new

The Hear layer led the week with 5 of 17 filings, from Soniox, Bland AI, Reson8, Gladia and 1 more.

Hearspeech-to-text5

SonioxModel providermodel release2026-06-11

Soniox ships v5 Async STT for structured speech data

Opens the v5 stack on batch audio five days before Real-Time — speaker-aware, language-aware transcripts aimed at automation, not cleaned-up walls of text.

Details2 sources

Soniox released stt-async-v5 for recorded audio, with higher multilingual accuracy across 60+ languages, reengineered speaker separation, native language ID, session context injection, and stronger alphanumeric formatting of numbers, emails, codes, and names.

Soniox v5 Async (stt-async-v5)

Bland AIPlatformlaunch2026-06-11

Bland noise cancellation V2 cuts word error rate 16% across 3,744 live calls

Bland builds its front-end filter itself rather than licensing Krisp's VIVA voice isolation (2026-10-07-krisp-voice-isolation-fp16); the 16% WER gain is measured on Bland's own traffic with no baseline WER stated, and it feeds the Fluent STT filed in April (2026-04-09-bland-fluent-stt).

Details1 source · 2 X posts

The filter removes background noise before Bland's transcriber hears it and can be switched on per persona, per call, or per inbound number.

Noise Cancellation V2

Tested across 3,744 live production calls: 16% more accurate transcription with noise cancellation enabled.

https://x.com/usebland/status/2065165259963539665

Introducing: Noise Cancellation v2

https://x.com/usebland/status/2065572240578474292
Reson8Model providerAPI2026-06-10

Reson8 releases a Turns endpoint for voice-agent speech-to-text

A 326 ms median time-to-final-transcript and in-model end-of-turn detection give voice-agent builders a dedicated streaming path, though the Pipecat-benchmark scores come from Reson8's own run.

Also Orchestrate

Details1 source

Reson8 released Turns, a streaming speech-to-text endpoint that predicts the end of a turn from the audio and language rather than a fixed silence timer, and flags barge-in without a separate interruption pipeline.

Turns endpoint

GladiaModel providermodel release2026-06-10

Gladia releases Solaria-3, a speech-to-text model tuned for noisy European business audio

A European STT vendor shipping a five-language model for contact-centre and meeting audio, and publishing where it loses, gives buyers a check against public leaderboards built on clean read speech.

Details3 sources

Solaria-3 is Gladia's pre-recorded STT model for English, French, German, Spanish, and Italian calls and meetings, sitting beside Solaria-1, which keeps the 100-plus language coverage and wins on clean read speech.

Solaria-3

DeepgramModel providermodel release2026-06-10

Deepgram ships Batch Diarization V2 for more accurate speaker labelling

Speaker attribution is the handoff between raw transcription and usable QA or clinical records; Deepgram's 3.3x preference result comes from its own human evaluation, not an independent benchmark.

Details2 sources · 1 X post

Deepgram's Batch Diarization V2 labels speakers in pre-recorded audio with improved attribution accuracy and a new diarize_model parameter, with no price or breaking API change.

Batch Diarization V2

Introducing Batch Diarization V2, a major upgrade to speaker labeling for pre-recorded audio. Understanding what was said is only part of the story. You also need to know who said it. Highlights: 🔹 Preferred 3.3X in human evaluation 🔹 Improved speaker attribution accuracy…

https://x.com/DeepgramAI/status/2064743195624227313

Speaktext-to-speech3

ZyphraModel providermodel release2026-06-12net-new

Zyphra releases ZONOS2, an 8-billion-parameter open MoE TTS model

Moving the Zonos line from a 1.6-billion-parameter beta to an 8-billion-parameter MoE design, with Zyphra reporting a four-times rise in real-time throughput, gives builders an Apache 2.0 alternative for high-fidelity cloning, though the naturalness claims and ZTTS1-Eval results are the company's own.

Details4 sources · 1 X post

ZONOS2 is a sparse mixture-of-experts text-to-speech model with 900 million active and 8 billion total parameters, Apache 2.0 weights, zero-shot voice cloning, and a hosted path on Zyphra Cloud.

ZONOS2

Today we're releasing ZONOS2, our next-generation real-time TTS model with high-fidelity voice cloning.

https://x.com/ZyphraAI/status/2065498869954490846
Resemble AIModel providermodel release2026-06-10

Resemble ships Chatterbox Multilingual v3 with default PerTh watermark

Open-weight multilingual TTS that watermarks every self-hosted sample by default — compliance feature, not an optional add-on, timed ahead of Article 50.

Details2 sources

Resemble released Chatterbox Multilingual v3, an open-weight 0.5B MIT-licensed TTS covering 25 languages with PerTh watermarking embedded by default, plus an NVIDIA NIM path and Single-Language Pack checkpoints for priority dialects.

Chatterbox Multilingual v3

GradiumModel providermodel release2026-06-10

Gradium upgrades default TTS for hard-case pronunciation

Default-path pronunciation fix for the phone-quality cases that break agents — emails and codes — weeks before the Aug hard-case eval packaging.

Details1 source

Gradium made a new TTS model the API default, targeting email, phone, acronym, and code read-out accuracy across English, French, Spanish, Portuguese, and German, with prior voices carrying over and no migration step.

Gradium TTS (June upgrade)

Showavatars, video1

ElevenLabsModel providerlaunch2026-06-11

ElevenLabs adds avatars to ElevenCreative for talking videos

Voice and avatar creation now share one ElevenCreative path, expanding ElevenLabs beyond audio output; the launch does not establish a new avatar model or benchmark.

Also Speak

Details2 sources · 1 X post

ElevenCreative can combine a script, an ElevenLabs voice, and an avatar to create studio-style talking videos in one workflow.

Avatars in ElevenCreative

Introducing Avatars in @ElevenCreative. The best AI voices, now with a face. Create studio-grade talking videos from a script, a voice, and an avatar - all in one place.

https://x.com/ElevenLabs/status/2065129866471551160

ThinkLLMs, speech-to-speech2

KrispPlatformlaunch2026-06-09

Krisp Voice Translation v3 adds per-call accuracy scoring and opens the engine as a self-serve API

Krisp's February SDK (2026-02-18-krisp-voice-translation-sdk) reached CX platforms; v3 adds the QA and audit trail regulated contact centres need and sells the engine directly to developers, against Konecta's V2V (2026-09-24-konecta-v2v-voice-translation). The 96% accuracy is Krisp's AutoQA score, not an independent test.

Also HearSpeak

Details3 sources

Contact centres get quality scoring on every translated call, live audits, and pre-translated Quick Phrases, while developers can call the same speech-to-speech translation engine through a self-serve API.

Krisp Voice Translation v3 and Voice Translation API

GoogleModel providermodel release2026-06-09

Google ships Gemini 3.5 Live Translate for continuous S2S

Puts continuous S2S translate into Gemini Live API, Meet, and consumer Translate on one day — 70+ languages with SynthID, and builder paths that name both Pipecat and LiveKit.

Also SpeakHear

Details1 source

Google released Gemini 3.5 Live Translate, a continuous speech-to-speech translation model covering 70+ languages with preserved intonation, SynthID watermarking, developer preview on the Gemini Live API and AI Studio, Meet private preview, and Google Translate app rollout.

Gemini 3.5 Live Translate

Orchestrateframeworks, evals4

Daily / PipecatPlatformlaunch2026-06-13

Pipecat Context Hub 0.2.0 publishes to PyPI and adds one-shot CLI commands for every tool

The CLI is the retrieval path the skills-only plugin filed in W42 (2026-10-15-pipecat-context-hub-plugin-0-2-3) calls, so this release is the packaging step that plugin relies on.

Details2 sources

Pipecat Context Hub 0.2.0 is the first PyPI release, runnable without a clone through uvx pipecat-ai-context-hub, and exposes every MCP tool as a one-shot CLI subcommand.

Pipecat Context Hub 0.2.0

LiveKitPlatformlaunch2026-06-11

LiveKit Agents 1.6.0 (Python + JS) introduces asynchronous tools

Major tool-runtime change for voice agents that call slow backends, filed once for Python and Node.

Details2 sources

livekit-agents and @livekit/agents 1.6.0 introduce asynchronous tools so long-running tool calls can stream progress updates and keep the conversation alive instead of silent waits.

livekit-agents / @livekit/agents 1.6.0

CrestaPlatformlaunch2026-06-11net-new

Cresta launches Conductor agent-building agent for CX

Turns Cresta's conversation data and Testing Suite into an agent that builds the CX agent — blueprint-before-code, not another prompt playground.

Also Use

Details2 sources

Cresta launched Cresta Conductor, a natural-language agentic engine that discovers requirements, produces a reviewable blueprint, generates prompts, subagents, configs, and custom code, then tests with Synthetic Customers and optimises from production transcripts.

Cresta Conductor

ParloaPlatformlaunch2026-06-08

Parloa launches Agent Skills, MCP-based tool chains that business teams configure without code

The four-to-eight-week onboarding claim is the sales case for MCP over bespoke APIs, and it rests on early results from one unnamed travel customer. Cresta shipped its own agent-building agent this week (2026-06-11-cresta-conductor); that is an analytical comparison, not a documented link.

Also Connect

Details1 source

Agent Skills build integration chains to CRMs, booking engines, and ticketing systems on the model context protocol, with deterministic tool calls and configurable Success Conditions for business outcomes.

Agent Skills

Connecttelephony, channels1

VobizPlatformlaunch2026-06-11

Vobiz ships Number Health Analytics for every account

Adds number-level campaign observability on the same telephony accounts, not a new call-control API — same calendar day as a LiveKit Agents filing, with no stated link between them.

Also Orchestrate

Details2 sources · 1 X post

Vobiz said Number Health Analytics is live on all accounts: a per-number dashboard of answer rates, failed-call trends, average duration, and a 30-day volume chart.

Number Health Analytics

Number Health Analytics is now live: a real-time dashboard that shows you answer rates, failed call trends, average duration of calls, and a 30-day daily volume chart for each number.

https://x.com/Vobiz_Ai/status/2064987100697723203

Useagents in market1

KrakenEnterprisepartnership2026-06-11net-new

Kraken launches utility customer-service agents with Sierra

A utility-specific data model supplies the guardrails generic customer-service agents lack for regulated energy journeys; the four-week launch and 1.3 million-account coverage are partner-reported.

Also Connect

Details2 sources · 1 X post

Kraken's Autonomous Agents combine Sierra's customer-experience platform with Kraken's utility data model and workflows for governed customer-service journeys, including voice.

Autonomous Agents · vendor Sierra

Four weeks from project launch to live production. That's how fast Kraken deployed utility-specific AI agents that handle complex, regulated, domain-specific journeys for 1.3 million customer accounts. Read more: https://t.co/C6i16XwQ6F

https://x.com/SierraPlatform/status/2065125765566669238