Search AI Products and News

Alibaba has fully launched the Qwen-Audio-3.0 series of speech models on the Tongyi AI platform, including three types: speech recognition, speech synthesis, and real-time voice interaction. Officially, the series performed exceptionally well in the July 2024 speech ranking list by the authoritative evaluation platform Artificial Analysis, securing first place in all three categories: speech recognition, real-time interaction, and speech synthesis.

NVIDIA has released its first open-source end-to-end full-duplex speech dialogue model, NemotronLabs VoiceChat 11B, which uses a single network to directly perform streaming speech understanding and generation, replacing the traditional ASR, LLM, TTS step-by-step architecture, achieving breakthroughs in real-time speech interaction.

Tencent Hunyuan launches the speech recognition model Hy ASR 3.0 preview, which leverages the language comprehension capabilities of the large language model Hy3, combining high-precision recognition with deep semantic understanding. It has been significantly improved in general recognition, context awareness, multi-scenario robustness, and dialect coverage. It can provide more accurate and coherent transcriptions by combining context, achieving an evolution from word-by-word transcription to semantic understanding.

Tencent Hunyuan launches Hy ASR3.0preview, integrating Hy3 LLM's semantic understanding into speech recognition for context-aware, direct transcription. On open benchmarks, word error rates reach 3.34% for Mandarin, 2.62% for English, and 3.12% for Cantonese, achieving around 3% overall with impressive multi-dimensional performance.....

Microsoft tests MAI Realtime, its first native real-time speech model. It supports 16 languages and 2 voice styles, enables simultaneous listening and speaking, allows seamless interruption, auto-detects and switches languages. Uses bidirectional full-duplex interaction, breaking turn-taking norms, with endpoint detection.....

Alibaba released the large speech recognition model Qwen-Audio-3.0-ASR-Flash, focusing on accurate identification of professional vocabulary. The model optimizes context consistency, industry-specific words, and hot word customization, supports speech polishing, and directly outputs structured text. The team has built a high-quality vocabulary database covering multiple industries such as medical, IT, and stock markets. Evaluation shows significant improvement in industry-specific vocabulary recognition performance.

On July 31, Alibaba's Tongyi Qianwen launched a speech recognition model featuring long audio context memory for consistent meeting transcription, and built-in industry lexicons (medical, IT) with 95.36% and 91.87% professional term recall, eliminating manual configuration. Available on Alibaba Cloud Bailian platform.....

xAI releases the Grok Voice Think Fast 2.0 speech recognition model for developers, achieving improvements in intelligent level, transcription accuracy, conversation ability, and tool calling efficiency. The pricing is $0.08 per minute. This model enhances performance without requiring adjustments to existing prompts. It achieved a total score of 82.9% in the Artificial Analysis benchmark test, surpassing the previous version's 75.7%.

Fish Audio launched the production-grade voice model S2.1 Pro for real-time conversations. Available via API with free quota, it offers ~90ms first-frame latency, supports 83 languages, and uses unified voice tech for smooth real-time dialogue.....

OpenRouter launches an audio transcription endpoint with a unified API key. Developers just need to send Base64 audio to directly get transcribed text, without needing to integrate Whisper or third-party SDKs separately, completely eliminating the sense of separation between chatting and transcription.

China's national security agency warns that AI voice cloning can replicate voices from just 3-5 seconds of audio, with ordinary people failing to detect fakes half the time. One-click generation tools on many platforms increase misuse risks, threatening social order and property safety.....

Qwen released Qwen-Audio-3.0-TTS, a text-to-speech model featuring free-style natural language instruction control. Users can simply describe desired tone, pace, and style in everyday language, eliminating the need to adjust engineering parameters and drastically lowering the usage barrier.....

Qwen introduces the real-time speech recognition model Fun-ASR-Realtime, reducing the first-word latency to the millisecond level and achieving a smooth "speak and feedback immediately" interaction. Its recognition accuracy is close to that of offline models, achieving high precision while breaking through the real-time performance bottleneck, marking a new height in voice interaction experience.

xAI launches the Voice Agent Builder beta version, allowing enterprises to build high-level speech AI agents in just two minutes through a no-code platform and its self-developed Grok Voice model. Its core is a highly integrated end-to-end architecture that addresses the pain points of traditional solutions with fragmented processes such as speech-to-text conversion, significantly lowering development and operational barriers.

Google released the Gemini 3.5 Live Translate audio model, achieving a breakthrough in cross-language communication through real-time speech-to-speech technology. It has been integrated into products such as Google AI Studio, Translate, and Meet. Its core innovation lies in pursuing naturalness in translation, abandoning the traditional turn-based translation model, with the aim of breaking down geographical and cultural barriers in language communication.

Global AI is undergoing an 'AI-native' technological shift. Addressing the current 'language-centric, externally patched vision or speech' architecture, a team released and open-sourced the native multimodal large model LongCat-Next and a discrete tokenizer, aiming to break modal barriers and enable AI to understand the physical world like processing text, achieved by reconstructing the underlying architecture.....

YouTube introduced a new podcast feature for Premium subscribers, aiming to compete with streaming giants and attract Spotify and Apple Podcast users. The key updates include personalized discovery and hands-free operation, enhancing the listening experience. The "Ask Music" feature has been expanded to podcasts, helping users avoid aimless searching and achieve smart recommendations.

On May 28, 2026, the global AI evaluation platform Artificial Analysis released a speech ranking list. Alibaba's Fun-Realtime-TTS-Preview speech large model obtained 1190 points, ranking fifth globally and first nationally. The model ranked first in three core speech fields such as ASR, demonstrating its comprehensive leading strength.

ElevenLabs introduces Stan Lee's iconic voice, professionally restored for its warmth and humor. Creators can synthesize speech via the Iconic Marketplace or use it in ElevenReader for reading articles, scripts, and applications in videos or podcasts.....

The speech generation model StepAudio2.5TTS from Chinese company StepXingchen has entered the top three globally in the Artificial Analysis Speech Arena Leaderboard, becoming the highest-ranked Chinese large model product on the list. The ranking uses a blind-test Elo scoring system, where users evaluate speech perception without knowing the model's identity, highlighting its genuine speech synthesis capabilities.