New Benchmark for Speech-to-Text: ElevenLabs Leads, Google Gemini Ranks Second with Comprehensive Capabilities

Artificial Analysis has recently released version 2.0 of its speech-to-text benchmark test (AA-WER v2.0). Test results show that ElevenLabs and Google demonstrate strong dominance in the field of audio transcription.

In terms of the core word error rate (WER) metric, ElevenLabs' Scribe v2 achieved an extremely low error rate of 2.3%, ranking first. Following closely behind is Google's Gemini3Pro, with an error rate of 2.9%. Notably, Google did not specifically train Gemini for transcription tasks, and this outstanding performance is entirely attributed to its strong multimodal general capabilities.

Other mainstream models performed as follows:

Mistral Voxtral Small: Ranked third with an error rate of 3.0%.
Google Gemini3Flash: Performed steadily, with an error rate of 3.1%.
OpenAI Whisper Large v3: As the most popular open-source model, it ranked in the middle with an error rate of 4.2%.
Bottom tier: Alibaba's Qwen3ASR Flash (5.9%), Amazon's Nova2Omni (6.0%), and Rev AI (6.1%) ranked at the bottom in the test.

In the specialized AA-AgentTalk test for voice assistant commands, the ranking remained stable. ElevenLabs' Scribe v2 and Google's Gemini3Pro led with error rates of 1.6% and 1.7%, respectively, demonstrating high reliability in handling short and direct voice interactions.

Ignoring Controversy! Meta Pushes the 'Name Tag' Feature for Ray-Ban Smart Glasses, Sparking a Discussion on Privacy and Ethics

Meta plans to introduce the 'Name Tag' feature for Ray-Ban smart glasses, which uses real-time facial recognition technology to allow wearers to identify others' identities and access social information. This move has sparked controversy in a society sensitive to privacy protection, as it is criticized for exploiting the public's desensitization to the normalization of surveillance, challenging the boundaries of privacy and technological ethics.

OpenRouter: Chinese AI Models Surpass US in Global Usage for the First Time, Domestic Forces Like MiniMax Dominate the Rankings

Chinese AI models have surpassed the United States in global usage for the first time. According to OpenRouter data, during the period from February 9 to 15, the weekly call volume of Chinese open-source models represented by MiniMax and Moonshot reached 410 million, marking a significant shift in the global AI competition landscape.

Free Yourself from Your Desk! Claude Code Launches Remote Control Feature: Control Local AI Programming with Your Phone

Anthropic has introduced a "remote control" mode for its AI programming assistant, Claude Code, allowing users to directly issue programming commands to a locally running AI agent via their phone. This feature connects desktop and mobile devices through a "secure synchronization layer," without requiring a cloud environment. Simply scan the terminal's QR code with your phone to remotely operate it, aiming to break physical office limitations and enhance development flexibility.

New Benchmark for Speech-to-Text: ElevenLabs Leads, Google Gemini Ranks Second with Comprehensive Capabilities

Related Recommendations

Ignoring Controversy! Meta Pushes the 'Name Tag' Feature for Ray-Ban Smart Glasses, Sparking a Discussion on Privacy and Ethics

Microsoft Reaffirms Core Partnership with OpenAI, Azure's Exclusive Position Unshakable

OpenRouter: Chinese AI Models Surpass US in Global Usage for the First Time, Domestic Forces Like MiniMax Dominate the Rankings

Free Yourself from Your Desk! Claude Code Launches Remote Control Feature: Control Local AI Programming with Your Phone

Google's Major Upgrade to AI Creative Studio Flow Integrates Image and Video Generation