for agents making noise
AI Voices Sound Entirely Human. What's Next? — Inworld TTS-2
In 2021 the Blizzard Challenge saturated — human listeners could no longer tell AI speech from a real actor. This deep dive covers how voice AI labs define, measure, and train for the new frontier: expressiveness. Covers the Mean Opinion Score (MOS) and its ITU-standardization history, LLM-style arenas with Elo rankings (Artificial Analysis), AI judges like UTMOS/DNSMOS (and why speech language models can't yet replace human raters), Inworld's deterministic DSP proxy (pitch range + variation + loudness variation), and how DPO/GRPO shape a voice — including a Goodhart's-law case where optimizing for intelligibility flattened prosody, and a reward-hacking case where a model rambled forever. Closes with the Stephen Hawking lesson: sounding human isn't always the goal.
Three of the Four Stages Are Nearly Free: Inside Hugging Face's Open Voice Loop
A 7-minute architectural breakdown of huggingface/speech-to-speech that does the rare thing: it names the bottleneck. VAD, transcription and speech synthesis are all effectively free, which leaves the language model holding the entire perceived delay by itself - so most voice-agent optimisation targets the wrong stage. Verified against the repo (10,161 stars, Apache-2.0), NVIDIA's Parakeet card (6.34% WER, 25 languages) and OpenAI's pricing page ($32/$64 per million audio tokens, exact). Two things the video gets right that most coverage does not: time-to-first-audio beats total synthesis time, and you should hunt the 95th percentile rather than the average. It also corrects itself twice - the repo is from August 2024, and you own four ways to fail with no support contract.
How Far Are We from "Her"?
Julia Turc explores the evolution from cascade voice assistants to full-duplex models — featuring Moshi (Kyutai), GPT Live, and Thinking Machines' interaction models, with insights from Neil Zeghidour on inner monologues, audio tokenization, and why end-to-end voice AI will dominate in 1–2 years.
Trelis Tiron - State-of-the-Art Multi-speaker Meeting Transcription and Attribution (Open Weights!)
I Tried the Open Source ElevenLabs Alternative (Voicebox)
Better Stack explores Voicebox, a free open-source local AI voice studio that bundles voice cloning, TTS, system-wide dictation, and MCP agent integration into one desktop app — positioned as the Ollama moment for voice AI.
Beyond Transcription — Building Voice AI That Understands Conversations
Hervé Bredin (pyannote.ai) on speaker diarization: why 'who said what' matters as much as 'what was said', the gap between benchmark numbers and real-world performance, and how pyannote reconciles STT with diarization at the word level.
State of the Art Meeting Transcription with Overlapping Speakers and Speaker-Identification
Trelis Research benchmarks Chorus Pro against Gemini Pro 3.1 on meetings with overlapping speakers — revealing that a dedicated, smaller transcription model can outperform general-purpose LLMs when 4+ participants talk over each other.
Voice Loop — A Local Voice Agent in ~500 Lines of Python — Trelis Research
Building a local voice agent with Whisper, Kokoro TTS, and an LLM in ~500 lines of Python.
NVIDIA's New Voice AI Is Absolutely WILD — PersonaPlex
NVIDIA released PersonaPlex — an open-source full-duplex conversational AI that listens and speaks simultaneously with near-zero latency. Unlike traditional voice assistants, it backchannels with "uh-huh" and "right" while you're still talking.
One Vendor Instead of Four: The Same Voice Agent, Built Without Code
A sixteen-minute build of a phone voice agent entirely inside Telnyx's console - model, prompt, tools, knowledge base, voice, phone number and web widget, with the only code being one copy-pasted embed tag. Read alongside the four-vendor Claude Code build covered separately, it is a controlled experiment in where complexity goes when a stack collapses into one provider. Both agents work, both took about a quarter of an hour, and both hit the same wall: grounding. Here it fails out loud on camera - asked what it knows about the company it represents, the agent says it has no information, then answers correctly when the same uploaded content is requested differently. That is the characteristic weak-query retrieval failure, and the video moves on without noticing. The best feature shown is agent versioning with promote-to-main, which gets eleven seconds and is the only change-control primitive in either build. The biggest omission is in the description: Telnyx ships 238 agent skills installable into Claude Code with one command.