Voice AI Weekly #4 - 1st Week of August 2026
This week, speed and signal understanding were the two prominent pillars in the voice AI industry. We have summarized five news items from overseas.
1. OpenAI Unveils Latency-Free Real-time Voice AI Architecture
OpenAI disclosed the internal structure of its 3rd-generation voice system, GPT-Live, on August 3rd. By eliminating turn detectors and processing voice in a full-duplex manner, it performs listening and speaking simultaneously. The session connection process has also been reduced from the previous 6 round-trips to just 1, significantly boosting response speeds.
Source How we built a realtime system for responsive voice AI in six months
2. PolyAI Launches Audio-Native Conversational Model, Dialog-RSN-1
On July 30th, PolyAI introduced Dialog-RSN-1, which bundles turn-taking, speech recognition, function calling, and response generation into one. It directly processes voice signals without going through a text conversion step, achieving a response speed of under 300 milliseconds. While the input is audio-native, the voice synthesis is separated into an independent TTS, retaining voice control.
This structure, which reads intent and emotion directly from the signal layer, shares the same direction as MAGO's Acoustic Intelligence. It appears that voice AI for contact centers is moving towards bypassing text entirely.
Source Dialog-RSN-1: a voice model that hears calls the way humans do
3. OVH Group Completes Acquisition of Voice AI Company Gladia
OVH Group finalized the acquisition of French voice AI startup Gladia on July 31st. Gladia is an STT platform that transcribes more than 100 languages in real-time, securing 300,000 developers and 2,000 corporate customers. With this acquisition, OVH has internalized voice technology into its own AI infrastructure, strengthening its sovereign cloud strategy.
The movement to process voice data within national infrastructure has become clear in Europe as well; this also serves as proof that more places are looking for on-premise voice AI.
Source OVH Groupe completes the acquisition of Gladia, expert in voice AI
4. Russian Research Team Captures Cognitive Decline Signals Solely Through Voice in Daily Conversations
A Russian IKBFU research team unveiled a neural network on July 29th that analyzes 88 acoustic features in ordinary conversation recordings to detect early signs of dementia and brain damage. It differs from previous research in that it works solely with natural conversation, without set test questions. The research team stated that they designed this system as a screening tool rather than a diagnostic tool.
Source Voice AI Flags Cognitive Decline From Ordinary Speech, Catching What Clinicians Miss
5. Hume AI, Benchmark Comparing Emotion Understanding of 40+ Voice AI Models Emerges
On July 15th, Hume AI released Real World VoiceEQ, a benchmark that evaluates the actual conversational quality of voice AI. Based on over 1 million human evaluations, it compared over 40 models across 15+ categories and 60+ metrics. Google Gemini scored the highest in text-to-speech, while ElevenLabs took the top spot in speech recognition.
Source Introducing Real World VoiceEQ: Measuring the human quality of voice AI
Response speed and signal understanding were the key issues in the voice technology sector this week.
Please feel free to contact contact@holamago.com for any questions or collaboration inquiries. Next week, we will bring you more voice AI news chosen from MAGO's perspective.