Voice AI Weekly #12 - 1st Week of October 2026

Here is this week's global Voice AI news. We have covered real-time STT competition, on-device models, dialect voice recognition, and investment news.

1. Microsoft Unveils 3 Real-Time Voice Models for Contact Centers

Microsoft unveiled three real-time voice models on October 1st. MAI-Transcribe-2-Streaming automatically detects 60 languages and transcribes them, delivering the first transcription results within hundreds of milliseconds. MAI-Voice-2.1 for voice generation and the lightweight MAI-Voice-2.1-Flash were also released. Microsoft is promoting its use for understanding customer requests before sentences are completed during a call. As the competition for speed and multilingual capabilities intensifies, the ability to read emotions and status signals that are not captured in text is expected to become a clearer differentiator.

Source PYMNTS

2. ElevenLabs Doubles Valuation to $22 Billion

ElevenLabs raised its valuation to $22 billion upon completing a $300 million employee stock sale (tender offer). Since it was valued at $11 billion in February, the valuation has doubled in less than nine months. The round was led by Wellington and T. Rowe Price, with new participation from EQT, Goldman Sachs, and others.

Source Tech.eu

3. A 164MB Open-Source STT, Phonon-2

Phonon-2, released by Fermion Research, has a size of 164MB. Its average Word Error Rate (WER) across 7 English datasets on the Open ASR Leaderboard is 5.21%, which is lower than Whisper large-v3-turbo's 6.58%. It is a model distilled from Nvidia's Parakeet TDT and runs at 174 times real-time speed on an M5 MacBook Air. The license is CC-BY-4.0, English only.

Source Hugging Face

4. Morocco Releases Dialect Voice Recognition Model with Mistral

Morocco's Ministry of Digital Transition, in partnership with Mistral AI, has announced open-source AI tools. This includes a language identifier to distinguish Moroccan Arabic, and Voxtral, a speech recognition model that transcribes dialect speech. This model also handles code-switching where Arabic, French, and English are mixed. The government aims to apply it to public administration, education, and customer support, while targeting language-centric AI sovereignty.

Source TechAfrica News

5. Vocca Raises $20 Million for Hospital Call-Responding Voice AI

Vocca, based in Paris and New York, has raised $20 million in Series A funding. The company provides a service where AI handles patient calls for hospitals and clinics, and is currently used by 15,000 medical professionals in the US and Europe. According to the article, it operates in 86 countries.

Source Sifted

Voice AI is becoming faster and smaller, and because of this, emotions in signals and domain accuracy are becoming the deciding factors.

If you are curious about Voice AI implementation or sentiment analysis, please feel free to reach out to contact@holamago.com. I will bring you a summary of the latest Voice AI news again next week.