ElevenLabs Unveils New Speech-to-Text Model

ElevenLabs, an AI startup that recently secured a whopping $180 million in funding, is famous for its audio-generation skills. But now, they've taken a bold step into new territory by launching their first stand-alone speech-to-text model, called Scribe.

Valued at $3.3 billion, ElevenLabs has been a go-to for many companies needing text-to-speech services, thanks to their huge collection of voices. Now, they're setting their sights on speech detection, aiming to take on big names like Gladia, Speechmatics, AssemblyAI, Deepgram, and OpenAI's Whisper models.

Scribe isn't messing around—it supports over 99 languages right out of the gate. ElevenLabs says it's got excellent accuracy for over 25 languages, with a word error rate of less than 5%. We're talking English (with a claimed accuracy of 97%), French, German, Hindi, Indonesian, Japanese, Kannada, Malayalam, Polish, Portuguese, Spanish, and Vietnamese, among others. The rest fall into different accuracy categories: high (5% to 10% word error rate), good (10% to 20% word error rate), and moderate (25% to 50%).

The company claims Scribe beats out Google Gemini 2.0 Flash and Whisper Large V3 in multiple languages, according to FLEURS & Common Voice benchmark tests.

ElevenLabs actually built the speech-to-text part for their AI conversational agent platform last year, but Scribe is their first go at a stand-alone speech detection model. In a chat with TechCrunch last month, CEO Mati Staniszewski spilled the beans on their plans to beef up their speech detection tech.

"We want to get better at understanding what you're saying in a conversation. We're not just about generating content anymore; we're moving into understanding and transcribing speech," Staniszewski said. "A lot of folks think speech-to-text is old news, but for many languages, it's still pretty rough. We think we can do better because we've got in-house teams to label data and give us quick feedback."

Scribe's got some cool features too, like smart speaker diarization to tell you who's talking, timestamps at the word level for spot-on subtitles, and auto-tagging of sound events like audience laughter. Plus, ElevenLabs is letting customers directly transcribe video content to add subtitles or captions in their studio.

Right now, Scribe only works with pre-recorded audio. But don't worry, the company says they're working on a low-latency real-time version soon. So, it's not quite ready for meeting transcriptions or voice note-taking just yet.

ElevenLabs is charging $0.40 per hour of transcribed audio for Scribe. It's a competitive price, but some rivals are offering cheaper rates for audio transcriptions, with a few different features thrown in.