hotwirednews Get GrokBot

Week 25

15–21 Jun 20262026-W25 · 29 filings · 23 companies · 8 net-new

The Hear and Orchestrate layers led the week, with 7 filings each.

Hearspeech-to-text7

Smallest AIModel providermodel release2026-06-18net-new

Smallest AI adds four East-Asian languages to Pulse STT Realtime

Four additional languages widen a realtime STT API’s APAC coverage in one release, although the only comparative result cited is Smallest AI’s own benchmark suite.

Details2 sources · 1 X post

Smallest AI added Japanese, Korean, Mandarin, and Cantonese to its Pulse STT Realtime model, initially for customers in the United States.

Pulse STT Realtime East-Asian language support

Today we're launching support for the following East-Asian languages in Pulse STT Realtime: 🇯🇵Japanese 🇰🇷Korean 🇨🇳Mandarin 🇭🇰Cantonese Streaming, production-ready, and outperforming Deepgram on our Asian-language benchmark suite. Currently available in the US. Building voice for APAC? Let's talk.

https://x.com/smallest_AI/status/2067597478929776757
SpeechmaticsModel providermodel release2026-06-17

Speechmatics launches Melia 1, a code-switching multilingual STT model for batch transcription

Gladia's Solaria-3 (2026-06-10-gladia-solaria-3, W24) narrowed to five European languages a week earlier, while Melia put all 55+ Speechmatics languages in one model; it was batch-only at launch, so Speechmatics' real-time voice-agent STT came separately as Linden 1 (2026-09-17-speechmatics-agent-stt, W38).

Details3 sources

Melia 1 turns recordings that move between any of Speechmatics' 55+ languages into one transcript with no language chosen in advance, in production preview on the Batch API, Portal, and SDKs.

Melia 1

DeepgramModel providerlaunch2026-06-17

Deepgram makes its Australia endpoint generally available

Sydney processing and storage give Australian healthcare, financial, legal, and public-sector buyers an onshore path without a new SDK integration; the LLM portion of a Voice Agent remains outside Deepgram’s residency guarantee.

Also Speak

Details2 sources · 1 X post

Deepgram opened a managed Australia endpoint in Sydney that keeps speech data storage and inference inside Australia while retaining the same APIs and models.

Australia endpoint

The Deepgram Australia endpoint is now generally available. 🇦🇺 https://t.co/mS6c4HNQGX runs in Sydney, with your audio, transcripts, and speech output processed and stored entirely within Australia. Same API and models. Moving over is a base URL change, no self-hosting required. Blog → https://t.co/IDRCmUoVX6

https://x.com/DeepgramAI/status/2067285167836696732
Alibaba / QwenModel providermodel release2026-06-17

Alibaba ships Fun-ASR-Flash on Model Studio, a file-transcription ASR model covering seven Chinese dialect groups and 30 languages

At $0.000035 per second (about $0.13 an hour in Singapore) with Wu, Min, Yue and other dialect coverage, Fun-ASR-Flash is the batch model that Alibaba's 9 July Fun-ASR-Realtime upgrade (2026-07-09-alibaba-fun-asr-realtime-upgrade, W28) is paired with, and it shipped the same day as Speechmatics' batch-only multilingual Melia 1 (2026-06-17-speechmatics-melia-1).

Details3 sources · 1 X post

Tongyi Lab's offline Fun-ASR-Flash snapshot fun-asr-flash-2026-06-15 transcribes audio of up to five minutes with context input, punctuation, and normalisation of numbers, dates, and amounts, and is billed per second of audio on Alibaba Cloud Model Studio.

Fun-ASR-Flash (fun-asr-flash-2026-06-15)

Our offline model, Fun-ASR-Flash, hit #1 worldwide on Artificial Analysis, achieving a 1.7% WER.

https://x.com/TONGYI_SpeechAI/status/2075094022075629729
SonioxModel providermodel release2026-06-16

Soniox ships v5 Real-Time STT with live speaker separation

Pairs the structured-speech v5 stack with a live path five days after Async — endpoint_sensitivity and speaker-aware streaming matter for agent turn-taking, not just WER.

Also Orchestrate

Details2 sources

Soniox released stt-rt-v5 for live audio, adding reinvented real-time speaker separation, spoken language ID across 60+ languages, streaming translation over 3,600 pairs, faster semantic endpointing with endpoint_sensitivity, and stronger alphanumeric formatting.

Soniox v5 Real-Time (stt-rt-v5)

Daily / PipecatPlatformlaunch2026-06-16

Pipecat 1.4.0 adds MoonshineSTTService

On-device Moonshine joins Whisper-family segmented STT options.

Details3 sources

pipecat-ai 1.4.0 adds MoonshineSTTService for Useful Sensors' Moonshine on-device ASR.

pipecat-ai 1.4.0

Shunya LabsPlatformlaunch2026-06-15

Shunya Labs launches multilingual Voice AI for 216+ languages

CPU-first Indic-depth stack dated 15 Jun targets the 90%+ India native-language preference claim on the wire — a regional alternative to English-first STT/TTS for regulated on-prem buyers.

Also SpeakUse

Details2 sources

Shunya Labs announced a multilingual Voice AI platform covering 216+ languages and dialects with deep Indic and South/Southeast Asian focus, CPU-first deployment for on-premise or cloud, aimed at BFSI, healthcare, government, and customer support.

Multilingual Voice AI platform

Speaktext-to-speech5

LiveKitPlatformlaunch2026-06-19

LiveKit Agents 1.6.2 adds Gemini 3.1 Flash TTS and Fish Audio prosody

Gemini and Fish Audio TTS updates.

Details1 source

livekit-agents 1.6.2 adds Gemini 3.1 Flash TTS streaming and Fish Audio speed/volume prosody params defaulting to s2.1-pro.

livekit-agents 1.6.2

AsyncModel providermodel release2026-06-18net-new

Async ships Async Pro v1.0 multilingual TTS on fal

A six-language TTS model with explicit pause, emphasis, and timing controls gives builders a hosted alternative for multilingual voiceover; fal’s page lists a commercial licence and per-character billing.

Details3 sources · 1 X post

Async Pro v1.0 became available on fal as a text-to-speech model covering English, German, French, Spanish, Italian, and Portuguese.

Async Pro v1.0 · vendor fal.ai

Async Pro v1.0 from @async_platform is now available on fal Multilingual text-to-speech across English, German, French, Spanish, Italian and Portuguese Natural pacing, clear pronunciation and smooth sentence transitions Advanced text normalization for mixed content https://t.co/C1JKj4k8my

https://x.com/fal/status/2067670531772453327
Maya ResearchModel providerfunding2026-06-17net-new

Maya Research raises $1.9M from South Park Commons

At $1.9M the round is small beside Sarvam's $234M Series B first close two days earlier (2026-06-15-sarvam-series-b), yet Maya later claims Hindi TTS rankings above Sarvam's Bulbul v3; the founder's post is the only primary found.

Details2 sources · 1 X post

The Indian voice-model lab closed a $1.9M round from South Park Commons, which it calls its first partners, to build conversational speech for people who use technology mainly by voice.

Excited to share that Maya Research has raised $1.9M from South Park Commons, our first partners in this journey.

https://x.com/Dheemanthredy/status/2067232200467308957
TelnyxPlatformAPI2026-06-16

Telnyx adds Inworld Realtime TTS 2 to its Voice AI platform

One Inworld voice path now covers Telnyx's managed assistants and custom REST/WebSocket pipelines; the release publishes no independent latency or quality benchmark.

Also Connect

Details2 sources · 1 X post

Telnyx made Inworld Realtime TTS 2 available through its TTS API and Voice AI Assistants for expressive speech in live conversations.

Inworld Realtime TTS 2 on Telnyx · vendor Inworld AI

Voice agents should not sound like they are reading from a script. Inworld AI's Realtime TTS 2 is now available on Telnyx, through the Voice AI platform and API. Developers can use expressive, low-latency voices for live conversations, with support for better tone, pacing, and multilingual delivery. See what’s new: https://t.co/Hi0pA2QT11

https://x.com/telnyx/status/2066991291498741819
Daily / PipecatPlatformAPI2026-06-16

Pipecat 1.4.0 adds AggregatedTextProgressFrame and will_be_spoken

Clients can render partial spoken text without guessing from raw TTSTextFrame bursts.

Details3 sources

pipecat-ai 1.4.0 adds AggregatedTextProgressFrame beside TTSTextFrame and will_be_spoken on AggregatedTextFrame, plus Azure private TTS endpoints.

pipecat-ai 1.4.0

ThinkLLMs, speech-to-speech2

Daily / PipecatPlatformAPI2026-06-16

Pipecat 1.4.0 adds realtime_service_mode turn contract

Stops cascade aggregators from fighting server-VAD realtime sessions.

Details3 sources

pipecat-ai 1.4.0 adds realtime_service_mode on LLMContextAggregatorPair with RealtimeServiceMetadataFrame for OpenAI Realtime, Grok, and Inworld locally-driven-turns examples.

pipecat-ai 1.4.0

InworldPlatformAPI2026-06-15

Inworld ships Mastra Realtime voice package for duplex agents

Gives Mastra builders a first-class duplex path on Inworld without hand-rolling cascade glue — tool calls stay mid-conversation on the same audio channel.

Also OrchestrateSpeak

Details2 sources

Inworld released @mastra/voice-inworld-realtime, binding its Realtime API WebSocket session to Mastra Agents so STT, LLM, expressive TTS, semantic VAD, barge-in, and tool calling run in one duplex stream.

@mastra/voice-inworld-realtime · vendor Mastra

Orchestrateframeworks, evals7

Retell AIPlatformlaunch2026-06-19net-new

Retell ships Colloquial Model, Expressive Mode, and live call ops

Bundles delivery polish (colloquial + expressive) with supervisor ops and native CRM on one platform week — reduces the custom glue most voice-agent teams still wire by hand.

Also SpeakConnect

Details1 source

Retell's 2026 launch-week changelog dated 19 Jun adds Colloquial Model (~50 ms phrasing rewrite), Expressive Mode with auto and manual emotion tags, Live Call Monitoring with whisper and takeover, Built-in CRM two-way sync to Salesforce and HubSpot, plus Custom Dashboards.

Colloquial Model, Expressive Mode, Live Call Monitoring, Built-in CRM

NewoPlatformlaunch2026-06-18net-new

Newo launches Vibe Mode for conversationally managing AI employees

Business users can retrain or roll back a live voice agent by chatting with it, with checkpoints as the guardrail; the press dateline says 9 June while the page displays 18 June, which the row uses.

Also Use

Details1 source

Newo launched Vibe Mode, which lets business users update, train, diagnose, and roll back voice AI employees through plain-English conversation instead of technical settings or support tickets.

Vibe Mode

LiveKitPlatformmodel release2026-06-17

LiveKit ships Turn Detector v1 with open eot-bench

Moves end-of-turn out of STT vendor lock-in: one Cloud-default detector across any STT/LLM/TTS stack, with a public bench so operators can compare cutoffs vs latency.

Also Hear

Details1 source

LiveKit released Turn Detector v1, an audio+semantic end-of-turn model that is now the default on LiveKit Cloud at no charge for hosted agents, plus open-weight v1-mini for CPU and the open eot-bench suite covering English and 13 other languages.

LiveKit Turn Detector v1.0 and eot-bench

Daily / PipecatPlatformlaunch2026-06-16

Pipecat 1.4.0 adds pipecat.evals behavioural eval harness

First-party behaviour evals become part of the release train instead of external harnesses only.

Details3 sources

pipecat-ai 1.4.0 ships pipecat.evals for scripted bot behaviour checks in text or audio mode, usable as a library and from the CLI.

pipecat-ai 1.4.0

Bland AIPlatformfunding2026-06-16

Bland raises another $50M, taking its funding past $100M

Bland's case to investors is owning every model in the call, which its 2026 filings bear out with Fluent STT (2026-04-09-bland-fluent-stt), noise cancellation V2 (2026-06-11-bland-noise-cancellation-v2), and Speech v3 (2026-08-04-bland-ai-bland-speech-v3); the post names no lead investor or valuation, and 3.5 million calls a week is a company figure.

Also SpeakHear

Details1 source · 1 X post

The round pays for Bland's in-house voice models and infrastructure for long, high-stakes phone calls in healthcare, financial services, and similar sectors.

Bland (Series C)

Sarvam AIPlatformfunding2026-06-15

Sarvam raises $234M in the first close of a $300M Series B led by HCLTech

An IT services group leading the round gives Sarvam's voice agents a route into HCLTech's enterprise accounts; the 2 million daily interactions and 500,000 monthly audio hours are company-reported and not broken out by customer, and the round's last $66 million had not closed at announcement.

Also HearSpeak

Details5 sources

Sarvam raised $234 million at a $1.5 billion post-money valuation in the first close of a planned $300 million Series B, with HCLTech committing $150 million as lead strategic investor and Bessemer Venture Partners joining Khosla Ventures and Peak XV Partners.

FinPlatformacquisition2026-06-15

Salesforce agrees to buy Fin, formerly Intercom, for about $3.6 billion

Fin Voice (2026-06-04-fin-voice-2, W23) and Agentforce Voice (2026-10-05-florida-prepaid-agentforce-voice) now share an owner; the deal closed on 10 September, ahead of the fiscal Q4 guidance, and Fin shipped Apex Flash 2.0 for Fin Voice under Salesforce a month later (2026-10-07-fin-apex-2-nemotron).

Details5 sources

Salesforce signed a definitive agreement to acquire the customer-agent company, whose Apex-powered AI Agent answers by phone as well as chat, email, WhatsApp, SMS, and Slack, to sit alongside Agentforce.

Fin AI Agent

Connecttelephony, channels3

Daily / PipecatPlatformAPI2026-06-19

Pipecat websocket-transport 1.7.0 accepts an optional auth token in connection params

Token-protected WebSocket bots can be reached from the browser client without hand-built URLs; 1.8.0 later fixed stale tokens stacking up on the stored URL across reconnects.

Details1 source

@pipecat-ai/websocket-transport 1.7.0 lets browser clients pass an optional auth token in the connection params.

@pipecat-ai/websocket-transport 1.7.0

TencentPlatformpartnership2026-06-16net-new

Tencent Cloud and Inworld bring Realtime TTS-2 into Tencent RTC

A 3,200-node RTC network and Inworld's vendor-claimed sub-130 ms first chunk put expressive TTS inside an enterprise communications path; the partnership is an integration, not evidence of a joint customer go-live.

Also Speak

Details2 sources · 1 X post

Tencent Cloud integrated Inworld's Realtime TTS-2 into Tencent RTC so developers can select expressive, low-latency voices in the RTC console and SDK.

Inworld Realtime TTS-2 in Tencent RTC · vendor Inworld AI

We’ve partnered with @tencentcloud to bring Realtime TTS-2 into @TencentGlobal RTC! This is a major step forward in bringing expressive, low-latency voice AI into enterprise agents across customer support, healthcare, financial services, and more. https://t.co/tss4JhRzal

https://x.com/inworld/status/2066899353600704550
Daily / PipecatPlatformAPI2026-06-16

Pipecat 1.4.0 adds websocket transport to the dev runner and bot-interrupted RTVI

Clients learn about barges without inferring from missing audio alone.

Details3 sources

pipecat-ai 1.4.0 adds websocket to the development runner -t choices and a bot-interrupted RTVI server message when output is cut off.

pipecat-ai 1.4.0

Useagents in market5

Virginia UrologyEnterprisepartnership2026-06-19net-new

Virginia Urology uses Relatient Dash Voice AI to cut call abandonment by 53%

This is a named customer deployment, not a Relatient platform launch; the source gives outcomes but no go-live date or independent validation.

Also HearConnect

Details3 sources · 1 X post

Virginia Urology's seven-location, 47-provider urology group implemented Relatient's Dash contact-center scheduling platform and AI voice agent to handle routine calls and capture clinical requests.

Dash Contact Center Scheduling and AI-powered Voice Agent · vendor Relatient

38% of calls resolved start to finish by AI

https://x.com/Relatient/status/2118780451527471152
Assort HealthPlatformlaunch2026-06-18

Assort Health launches Continuous Patient Conversations for connected patient engagement

A longitudinal context layer can carry a patient’s preferred language and scheduling pattern across five interaction surfaces, but Assort gives no independent outcome data for the launch.

Also Connect

Details2 sources · 1 X post

Assort Health introduced a patient-engagement layer that shares Patient Journey Memory across voice, chat, two-way texting, intake, and proactive outreach.

Continuous Patient Conversations

Patients don't experience healthcare as channels. They experience it as a relationship. But most practices are still operating with disconnected systems: The phone doesn't know what happened over text. The chatbot doesn't know what happened on the phone. The intake form doesn't know what happened anywhere else. So patients repeat themselves. Staff piece together context manually. And every interaction starts over. Today we're introducing Continuous Patient Conversations. A new foundation for patient engagement built around one shared patient context: Patient Journey Memory. Voice, chat, two-way texting, intake, and proactive outreach all share it, so every interaction builds on the last instead of starting from scratch. Because patients shouldn't have to tell their story twice. The result is a more connected experience for patients and less manual work for staff. And fewer patients slipping through the cracks. See how it all comes together → https://t.co/mXF3B8RQaq

https://x.com/assort_health/status/2067652330317656090
ParloaPlatformpartnership2026-06-17

Parloa becomes the first agentic AI provider on Alvaria's outbound platform

This takes Parloa from inbound service into regulated outbound dialling. The 'first and only' claim is the vendors' own, and the release does not describe consent, calling-window, or opt-out handling beyond naming Alvaria's orchestration role.

Also Connect

Details1 source

Parloa's voice agents will make compliant outbound calls through Alvaria's pacing, list management, multi-channel sequencing, and capacity controls.

Parloa on the Alvaria Intelligence Platform

KonectaEnablerlaunch2026-06-16net-new

Konecta launches Kolibri agentic CX orchestration platform

A BPO-scale enabler productising voice and digital agents with named CX partners (ElevenLabs, Uniphore, NiCE) and a live Telepizza voicebot reference — production ops, not another pilot deck.

Also OrchestrateConnect

Details1 source

Konecta launched Kolibri, an agentic AI orchestration platform with up to 80% pre-built CX use cases, ISO 42001 governance, FinOps dashboards, and open integrations including ElevenLabs, Uniphore, NiCE, Google Cloud, CrewAI, and Salesforce.

Kolibri

LuaPlatformlaunch2026-06-15

Lua launches US Home Services voice agents for trades

Verticalises Lua Voice for US trades with calendar booking and field-service CRM hooks — the 1-in-4 / 85% missed-call figures are vendor-stated industry framing.

Also Connect

Details1 source

Lua launched Lua for Home Services in the US: voice agents for HVAC, plumbing, roofing and related trades that answer calls, qualify leads, quote from a price book, and book jobs into calendars or field-service tools.

Lua for Home Services