Speech, Voice and Realtime Audio Models — 2026-10-29
Nvidia released a new open-source speech-to-text model supporting 40 languages, while NTT announced a voice dialogue AI with 0.43-second response times for call centers. Meanwhile, KT launched "Token Factory," a platform managing over 500 AI models to optimize token efficiency, and ElevenLabs introduced "Eleven Realtime" models with sub-600ms latency.
Speech, Voice and Realtime Audio Models — 2026-10-29
Top developments
Nvidia releases open transcription model for 40 languages
Nvidia released a free speech-to-text model capable of transcribing up to 40 languages. However, the model is currently limited to recordings rather than live audio, positioning it as a significant tool for meeting and medical transcription while leaving real-time applications to competitors like ElevenLabs.
NTT develops voice AI with 0.43-second response time
NTT announced the development of a voice dialogue AI technology designed for call centers, achieving an average response time of 0.43 seconds. The technology enables natural conversations by automatically retrieving necessary information and speaking according to instructions, aiming to improve service quality and reduce operator workload. NTT plans to demonstrate this at the "NTT R&D FORUM 2026" and is planning joint experiments for real-world application.
ElevenLabs launches "Eleven Realtime" speech-to-speech models
ElevenLabs introduced "Eleven Realtime," a new family of speech-to-speech models for voice agents. The models, "v1 Mini" for natural conversation and "v1 Max" for complex tasks, feature end-to-end tuning of speech recognition, turn-taking, LLM, and voice. They achieve time-to-first-audio latencies of 620ms and 570ms respectively, and are compatible with the OpenAI Realtime API.
Ant Group open-sources Ming-3.0 speech models
Ant Group released Ming-3.0, a suite of five MIT-licensed speech models covering text-to-speech (TTS), audio understanding, and recognition. This open-source release adds to the growing ecosystem of accessible voice AI tools, alongside other recent releases like Gradium’s 31M-parameter on-device TTS model "Phonon".
Local view
KT launches "Token Factory" to manage 500+ AI models
KT launched "Token Factory," a platform that allows enterprises to manage and select from over 500 domestic and international AI models. The platform features an auto-model router to select optimal models and uses Korean prompt compression technology to improve token efficiency by up to 25%. This move positions KT to monetize AI usage as a utility, similar to voice and data services.
Google DeepMind releases "Genie 4" with real-time voice simulation
Google DeepMind announced "Genie 4," a world simulation model capable of generating real-time interactive environments with native voice features, including music, environmental sounds, and dialogue. Users can modify the virtual world with voice commands, such as requesting "fire and arrows from the castle walls," which the AI then implements in real-time.
Context & numbers
ElevenLabs API Pricing and Latency
As of October 2026, ElevenLabs API pricing for Text-to-Speech is listed at $0.08 per 1,000 characters for multilingual models and $0.04 for Flash/Turbo models. Speech-to-Text pricing is $0.22 per hour for Scribe and $0.39 per hour for Scribe Realtime. The market has settled into three lanes: expressive models for offline content, low-latency models for real-time agents, and low-cost models for bulk processing.
Voice AI Scams and Safety
Recent reports highlight the risks of AI voice cloning, with scammers using cloned voices to fake kidnappings. Experts recommend families establish "safe words" to verify identities. The FBI has reported $893 million lost in AI-related scams, with voice cloning becoming a primary vector due to the ability to create convincing clones from just 3 seconds of audio.
On the radar
- NTT R&D FORUM 2026: NTT will demonstrate its new 0.43-second response voice AI at this upcoming forum, with plans for joint experiments in call center environments.
- KT Token Factory Rollout: Following the October 27 launch, KT plans to expand the "Token Factory" platform beyond enterprise customers to broader markets in the coming year.
- Open TTS Leaderboard: Hugging Face continues to update the Open TTS Leaderboard, which now evaluates over 8,000 TTS models, with recent leaders including Kokoro-82M, Supertonic-3, and Fish Audio S2 Pro.
This content was collected, curated, and summarized entirely by AI — including how and what to gather. It may contain inaccuracies. Crew does not guarantee the accuracy of any information presented here. Always verify facts on your own before acting on them. Crew assumes no legal liability for any consequences arising from reliance on this content.
Powered by
