hotwirednews Get GrokBot

Week 12

16–22 Mar 20262026-W12 · 16 filings · 10 companies · 4 net-new

The Hear, Orchestrate and Connect layers led the week, with 4 filings each.

Hearspeech-to-text4

Reson8Model providerfunding2026-03-19net-new

Reson8 raises a €5 million pre-seed led by Balderton

The €5 million pre-seed gives an Amsterdam speech lab capital for Europe-hosted GPUs and real-time customisation, months before Resonant-1 shipped on 8 April.

Details2 sources · 1 X post

Amsterdam's Reson8 closed a €5 million pre-seed led by Balderton Capital, with participation from NP-Hard Ventures, to build customisable European speech recognition on infrastructure it owns.

Pre-seed funding

We raised €5M to build speech recognition that customizes in seconds.

https://x.com/reson8_official/status/2034610994933125608
ModulateModel providerlaunch2026-03-18net-new

Modulate launches Velma Transcribe for conversational speech-to-text at lower cost

At approximately $0.03 per hour in Modulate's 18 March comparison, Velma targets call-centre and agent workloads where overlapping speakers and interruptions make flat transcription expensive; the cost and benchmark comparisons are vendor claims.

Details2 sources · 1 X post

Modulate launched Velma Transcribe, a low-latency speech-to-text API for real-world conversations at approximately $0.03 per hour of audio.

Velma Transcribe

ArgmaxPlatformlaunch2026-03-18

Argmax brings real-time Parakeet transcription to Android with its Pro SDK

A Kotlin-first SDK extends Argmax's on-device transcription from Apple to Android hardware, while the launch version supports real-time transcription only and depends on Google LiteRT for the wider device path.

Also Orchestrate

Details3 sources · 1 X post

Argmax Pro SDK for Android is a Kotlin-first SDK for real-time speech recognition with Nvidia Parakeet and Google LiteRT acceleration across Android devices.

Argmax Pro SDK for Android · vendor Google LiteRT

Introducing Real-time Transcription with Nvidia Parakeet on Android!

https://x.com/argmax/status/2033980792838246784
LiveKitPlatformlaunch2026-03-16

LiveKit Agents 1.4.6 adds AssemblyAI speaker diarization

Diarization is the hear headline for 1.4.6.

Details1 source

livekit-agents 1.4.6 adds speaker diarization support on AssemblyAI STT.

livekit-agents 1.4.6

Speaktext-to-speech2

AmazonModel providerAPI2026-03-20

Amazon Polly adds Bidirectional Streaming API and 10 Generative voices

HTTP/2 duplex TTS lets LLM token streams speak without sentence buffering — AWS's own 39% faster bench is vendor-run, but the API shape matches cascaded voice-agent stacks.

Details2 sources

Amazon announced general availability of a Bidirectional Streaming API for Polly's Generative engine, plus 10 new Generative voices across eight locales and Generative-engine availability in Europe (London) and Canada (Central).

Polly Generative TTS Bidirectional Streaming

LiveKitPlatformlaunch2026-03-16

LiveKit Agents 1.4.6 adds Murf AI TTS plugin

New TTS vendor plus speak-path text transforms.

Details1 source

livekit-agents 1.4.6 adds a Murf AI TTS plugin and callable TextTransforms with a built-in replace transform.

livekit-agents 1.4.6

ThinkLLMs, speech-to-speech1

PolyAIModel providermodel release2026-03-19

PolyAI releases Raven 3.5, a post-trained voice-agent LLM covering 23 languages

Raven is the in-house model behind PolyAI's agents, and 3.5 adds auto-reasoning while keeping the sub-300 ms latency; the wins over GPT-5 and Claude Sonnet 4.6 come from PolyAI's own benchmarks.

Details1 source

Raven 3.5 uses more training data and a GRPO and DPO post-training mix after an SFT warm-up for PolyAI's voice-agent LLM.

Raven 3.5

Orchestrateframeworks, evals4

LiveKitPlatformlaunch2026-03-19

LiveKit Agents 1.5.0 (Python + JS) enables preemptive generation and ships Agent.create on Node

Defining turn-taking defaults for the 1.5 line, with a Node functional-API surface in the same version. Class-2: folded quiet patch(es) 1.5.4, 1.5.8 into this row.

Details4 sources

livekit-agents 1.5.0 turns on preemptive generation by default with adaptive interruption so LLM/TTS can start before end-of-turn; @livekit/agents 1.5.0 adds a functional Agent.create() / AgentTask.create() API on Node.

livekit-agents / @livekit/agents 1.5.0

Daily / PipecatPlatformAPI2026-03-18

Pipecat Context Hub 0.0.8 adds call-graph filters to search_api

Frame-flow questions become index queries instead of reads of the virtual-environment source; the notes give no accuracy measurement for the extraction.

Details1 source

Pipecat Context Hub 0.0.8 extracts yields and calls metadata from method bodies and exposes both as search_api filters.

Pipecat Context Hub 0.0.8

Daily / PipecatPlatformlaunch2026-03-18

Pipecat 0.0.106 adds wake-phrase user turn start strategy

Always-on mics can gate turns on a phrase without a separate wake-word microservice.

Details3 sources

pipecat-ai 0.0.106 adds WakePhraseUserTurnStartStrategy with single_activation mode, plus helpers to compose default start/stop strategy lists.

pipecat-ai 0.0.106

DeepgramPlatformpartnership2026-03-17

Deepgram, Penguin Solutions, and Dell deploy a production-ready voice AI infrastructure stack

The post links Deepgram's STT, TTS, and agent stack to an enterprise infrastructure deployment with Penguin Solutions and Dell, but publishes no architecture details or independent performance result.

Details1 source · 1 X post

Deepgram says it partnered with Penguin Solutions and Dell Technologies to deploy optimised infrastructure for high-volume, real-time voice AI workloads.

Deepgram voice AI infrastructure reference architecture · vendor Penguin Solutions; Dell Technologies

Voice AI is only as strong as the infrastructure behind it.

https://x.com/DeepgramAI/status/2033974145642283021

Connecttelephony, channels4

Daily / PipecatPlatformAPI2026-03-18

Pipecat 0.0.106 adds Daily DTMF input and SIP dial-out helpers

IVR-style digit collection and SIP dial-out configuration land on the Daily path together.

Details3 sources

pipecat-ai 0.0.106 adds Daily DTMF input via on_dtmf_event plus sip_provider and room_geo on the Daily runner configure() helper.

pipecat-ai 0.0.106

Amazon ConnectPlatformlaunch2026-03-18net-new

Amazon Connect brings agentic S2S voice to London with three new voices

UK/EU contact-centre buyers get Connect S2S in London with en-GB and es-US voices on 18 March — a geo and locale unlock, not a new duplex model.

Also Think

Details1 source

Amazon Connect extended agentic speech-to-speech voice experiences to Europe (London) and added three S2S voices: Pedro (es-US), Amy (en-GB), and Brian (en-GB).

Agentic speech-to-speech (London + voices)

Amazon ConnectPlatformlaunch2026-03-17net-new

Amazon Connect voice AI agents add 13 languages to reach 40 locales

Connect voice agents jump from prior coverage to 40 locales on 17 March — useful for EU and MENA contact centres, with quality still AWS-claimed only.

Also Use

Details1 source

Amazon Connect added 13 language locales for voice AI agents, bringing the total to 40, covering Arabic (Saudi Arabia), Czech, Danish, Dutch (Belgium), English (Ireland, New Zealand, Wales), German (Switzerland), Icelandic, Romanian, Spanish (Mexico), Turkish, and Welsh.

Voice AI agents multilingual expansion

LiveKitPlatformlaunch2026-03-16

LiveKit Agents 1.4.6 adds trunk_config on WarmTransferTask

SIP trunk wiring for warm transfers.

Details1 source

livekit-agents 1.4.6 adds trunk_config to WarmTransferTask for SIP endpoint transfers.

livekit-agents 1.4.6

Useagents in market1

DaptaPlatformlaunch2026-03-19net-new

Dapta Campaigns runs scheduled AI outbound call lists with retries

Retry rules and calling windows become no-code settings for small-business owners seven months before Dapta's seed round (2026-10-13-dapta-seed-6m); the post says nothing about consent or calling-hour law, which anyone dialling a purchased list would need to check.

Also Orchestrate

Details1 source

Dapta users upload a CSV or Excel contact list, set calling days, time windows, and retry limits, and their existing Dapta voice agent works through the list and logs every call.

Dapta Campaigns