Meta's First Real-Time Audio Perception Model Arrives: Transcribe 20 People Speaking Simultaneously, $3 per 1000 Minutes

Meta has launched Muse Voice Transcribe, its first real-time audio-aware model. Developers can call this capability through the Meta Model API, with a pricing of $3 (approximately 20.2 RMB) per 1000 minutes of audio, equivalent to $0.18 per hour.

Speak and Type, Automatically Separate 20+ Speakers

The model integrates streaming automatic speech recognition, speaker separation, and endpoint detection in the same process. "Streaming" means that it transcribes as people speak, without waiting for the entire audio to finish before producing results. Instead, it continuously outputs text in small segments, reducing latency and making it suitable for real-time transcription, meeting notes, and call subtitles. When users speak, the system can output text in real time, without needing to wait until the end of the recording to process it separately.

Muse Voice Transcribe can separate different speakers in recordings with more than 20 speakers, and it can also identify when a speaker stops talking, thus automatically determining the boundaries of an input. In terms of language support, the model covers over 70 languages, with 25 languages verified at launch; it supports audio longer than one hour and natural language switching within sentences or between sentences. Developers can also improve recognition accuracy for specific languages, terminology, or scenarios by using language, keyword, and context bias.

Adaptive Delay Mechanism Balances Speed and Accuracy

To balance real-time response and recognition accuracy, Meta introduced a dynamic decision-making mechanism called "adaptive delay." The system does not make a fixed trade-off between speed and accuracy for all words, but instead determines how much additional audio is needed before outputting the recognition result for each word: for easier-to-recognize speech, the system will submit the transcription results faster; for words that are unclear, have many technical terms, or are in complex contexts, it will use more audio information from the surrounding context to improve accuracy.

According to Meta, as of September 1, 2026, Muse Voice Transcribe ranked first on the Artificial Analysis Streaming Speech-to-Text Leaderboard.

No Design Needed, One Sentence Creates a Picture! Google Launches the Strongest Canva Rival Google Pics

Google announced its entry into the creative design field by launching an AI image creation and editing tool called Google Pics, which will be deeply integrated into the Google Workspace ecosystem for enterprise and advanced users. It targets daily office and lifestyle design scenarios. Its core is the Nano Banana image generation model. Unlike traditional design platforms, it focuses on directly generating and editing images through prompts.

Reddit Masterpiece Makes Claude Interact with You Through Handwriting in Real Time, Providing an Immersive Learning Experience That Is Simply Outstanding

Recently, a Reddit developer created an innovative handwriting interactive learning software that allows AI Claude to engage in real-time "handwriting dialogue" with users on a tablet or drawing board, similar to the magical diary in Harry Potter. Users can simply write notes or ideas on a blank page, and Claude will recognize and respond, enabling bidirectional handwriting conversation, making digital learning more warm and magical.

OpenAI CEO Says He Doesn't Like Smart Glasses: It's Uncomfortable to Talk to Someone Wearing a Camera and Lights, Zuckerberg Temporarily Relieves Pressure

OpenAI CEO Altman stated clearly during an interview that he doesn't like smart glasses, saying it's uncomfortable to talk to someone wearing cameras and lighting devices, implying that OpenAI will not enter this field for now, so Zuckerberg doesn't need to worry about competition. His attitude had already been evident earlier, while Meta's smart glasses business is growing rapidly.

57-Year-Old Man Followed AI Advice to Take a Shortcut and Got Stranded in the Mountains, Exhausted, and Later the AI Laughed at Him. It's True!

Zhejiang fire department disclosed that a 57-year-old Hunan man surnamed Yang got lost hiking in Ningbo after blindly trusting an AI-recommended descent route; exhausted, he was rescued after calling police. Firefighters remind hikers not to blindly trust AI and to choose established routes for safety.....

Qwen Learning Function Upgrade: New Textbook-Synchronized Essay Tutoring, Interactive Explanations Expanded to High School and University Math, Physics, and Chemistry, All Free

On Sept 1, Alibaba's Tongyi Qianwen app upgraded learning features: free unit essay guidance synced to textbooks, junior high Chinese/math lectures, and photo-based problem solving extended to high school/college STEM. Essay coaching asks guiding questions instead of model essays, helping students structure ideas, e.g., 'My Beloved Thing'.....