Concept 1: What is Scribe?
Scribe is ElevenLabs' first Speech to Text (STT) model.
Think of it as a tool that listens to audio and converts it into written text โ like a highly accurate digital transcriptionist.
๐ฏ Key claim: It is described as the world's most accurate transcription model.
Concept 2: What is a Word Error Rate (WER)?
Before understanding why Scribe is impressive, you need to understand Word Error Rate (WER).
- WER measures how many words a model gets wrong during transcription
- Lower WER = Better accuracy
- Example: A 40% WER means 4 out of every 10 words are transcribed incorrectly
Scribe's WER Performance:
| Language | Accuracy |
|---|---|
| Italian | 98.7% accurate |
| English | 96.7% accurate |
| Serbian, Cantonese, Malayalam | Competing models exceed 40% WER; Scribe dramatically reduces this |
Concept 3: Benchmark Testing
Scribe was tested against industry-leading models:
- Gemini 2.0 Flash
- Whisper Large V3
- Deepgram Nova-3
Testing was done on two standard benchmarks:
- FLEURS โ tests across 102 languages
- Common Voice โ tests across 102 languages
โ Scribe outperformed all competitors on both benchmarks across 99 languages
Concept 4: Key Features of Scribe
Scribe isn't just about accuracy. It comes packed with three core features:
๐ 1. Word-Level Timestamps
- Marks exactly when each word was spoken in the audio
- Useful for: subtitles, searchable transcripts, video editing
๐ฅ 2. Speaker Diarization
- Identifies and separates different speakers in the audio
- Example: In a meeting, it can label "Speaker 1 said X, Speaker 2 said Y"
๐ 3. Audio-Event Tagging
- Detects non-speech sounds and labels them
- Example:
[laughter],[applause],[music]
Concept 5: Language Accessibility
One of Scribe's most important contributions is reducing the gap for underserved languages.
Many STT models are trained heavily on English and a few major languages. This leaves speakers of languages like:
- Serbian
- Cantonese
- Malayalam
...with very poor transcription quality (40%+ error rates).
๐ Scribe supports 99 languages and dramatically improves accuracy for these traditionally underserved communities.
Concept 6: How Can You Use Scribe?
There are two ways to access Scribe:
๐จโ๐ป For Developers โ API Integration
- Use the Speech to Text API
- Receive structured JSON responses containing:
- Full transcript
- Speaker diarization labels
- Word-level timestamps
- Non-speech event markers
- A low-latency version for real-time applications is coming soon
๐ฅ๏ธ For Creators & Businesses โ Dashboard
- Upload audio or video files directly
- Get formatted transcripts without writing any code
Summary Table
| Concept | Key Takeaway |
|---|---|
| What is Scribe? | ElevenLabs' Speech-to-Text model |
| Word Error Rate | Lower = better; Scribe achieves industry-low WER |
| Benchmarks | Outperforms Gemini, Whisper, Deepgram |
| Core Features | Timestamps, Speaker ID, Audio-event tagging |
| Language Access | Supports 99 languages including underserved ones |
| How to Use | API (developers) or Dashboard (creators/businesses) |
๐ก Bottom Line: Scribe is a highly accurate, feature-rich, and inclusive speech-to-text model designed to work reliably across languages, speakers, and real-world audio conditions.