Topic
29 cards
Brief
Block says Dessa was trained entirely on speech with no text data and runs locally on a MacBook Pro M4 Max. It is not public, and access is by sign-up form for security researchers.
Act on this
Try the Q4_K_M quant if you run local voice agents or game characters on an RTX GPU.
Act on this
Check the Voice Agents SDK pricing if you build voice agents on Mistral.
Brief
The Hacker News tutorial covers auditing voice agents with ZizkaDB and LiveKit. The full text was blocked, so no versions or dates were verified.
Brief
The models launched October 20 through ElevenAgents and an OpenAI-compatible Realtime API. Max reports 98.8 percent on Big Bench Audio and 0.57 seconds response time, but these figures are unverified.
Brief
The arXiv paper keeps in-flight tool calls alive across speech updates. It reports 16 more points of tool coverage and 36.4 percent lower final-response latency on FDBench v3.
Brief
The free TestFlight app runs speech recognition on-device and records no audio. It needs an Apple silicon Mac with the Claude desktop app and an iPhone on iOS 18 or later.
Hype watch
The Reddit and X demo claims under one second latency for GTA V NPCs. No public repo or release is confirmed.
Act on this
Build the open-source macOS menu bar app with full Xcode if you want dictation without a subscription.
Brief
Sipgate runs KugelAudio's speech model in German data centers for its Sona AI telephony. Available since September 22, but no certifications were given.
Brief
A note.com digest covers GMO Flatt Security, TOKIUM MCP support, GMO Sign, Kiva and neoAI. The core launch claims were checked and supported.
Brief
The Product Hunt launch converts live speech into translated audio. No version, pricing or API availability is stated.
Hype watch
A PR Newswire release claiming a voice interface can be added to any app in about an hour. The market claims are vendor-sourced.
Hype watch
The launch claims a full voice pipeline for $0.04 per minute and beats Whisper large-v3 on multilingual transcription. The cost and accuracy figures are vendor-stated.
Brief
A Medium comparison published October 21 whose full text was blocked. Pricing, quality and availability claims could not be verified.
Act on this
Try the Apache-2.0 early preview if your local assistant needs speaker-only audio before STT.
Hype watch
The announcement promises conversations that turn into revenue. The cached text is a truncated press teaser with no launch date or availability.
Act on this
Compare AssemblyAI's $4.50 an hour bundle with per-minute stacks before you pick a voice agent stack.
Act on this
Install the MIT-licensed Omarchy plugin if you run agent CLIs on Omarchy.
Act on this
Test Eleven Realtime v1 Max or Mini in ElevenAgents or the OpenAI-compatible Realtime API.
Brief
Adds speech-to-text, a TTS engine and an SDK for voice agents. The low-cost pricing could not be verified because the source page was blocked.
Brief
Uses your own Claude, ChatGPT, API key or local model. It needs Apple silicon and macOS 15 or later, and it asks before sending outbound messages.
Brief
A newsletters index built with Codex voice mode. It adds a Django model, a migration, and imports from Substack RSS.
Act on this
Test SPEV for voice agent turn detection, since the hosted API is free through October 2026.
Brief
A product announcement. Voice requests become tasks that the user must approve before they run. No engineering details are given.
Act on this
Try the macOS-only Claude Code plugin that speaks replies through a Stop hook.
Act on this
Test voice agents with background TV or broadcast audio before launch.
Brief
Launched 14 Oct via the Voice Agent API at $0.08 per minute. Vendor-reported benchmarks put it at 86.9 on the Artificial Analysis index.
Brief
Tool lookups start once key details are spoken, not after the caller finishes. SHANKS completed 56.9% of tool calls before turn end.
