Transcription compared
How accurate is scryp at transcription in your language – and how secure is your data compared to the big cloud services? Here are our benchmark results.
Transcription accuracy for your language
Word Error Rate (WER) – lower is better. Accuracy = 100% − WER.
← Scroll table sideways →
| Provider | General (accuracy) | Specialized domain (accuracy) | WER general |
|---|---|---|---|
| scrypSX-3 | 97.2 % | 96.1 % | 2.8 % |
| OpenAIWhisper Large v3 | 91.8 % | 83.5 % | 8.2 % |
| Google STTCloud Speech-to-Text | 88.5 % | 80.2 % | 11.5 % |
| Azure SpeechMicrosoft Cognitive | 87.9 % | 78.7 % | 12.1 % |
| AWS TranscribeAmazon Standard | 85.7 % | 75.9 % | 14.3 % |
Test dataset & methodology
- The basis is our own test dataset of publicly available audio recordings in your market across the General, Health and Legal domains.
- All providers were tested with identical audio files and a uniform evaluation method (Word Error Rate after standard normalization).
- The Specialized domain column shows the average across health and legal recordings. Error rates rise with specialized vocabulary for all providers – but far more for models without domain-specific training.
- Results may vary depending on audio quality, number of speakers, dialect and the specific use case.
Why scryp is built differently for sensitive content
These points describe scryp's own architecture. They're deliberately precise and not meant as a blanket statement about every other provider.
Encryption on your device
Files are encrypted in the browser before upload. Stored content is permanently kept encrypted only.
Clear EU architecture
Transcription in Austria, encrypted storage in Germany – both within the EU. That makes it transparent where each processing step takes place.
Our own processing infrastructure
No external third-party AI is involved in transcription. That reduces additional data flows and dependencies.
Productive browser workflow
Editing, exporting, sharing and audio sync are built right into the product, not just available as separate API building blocks.
Sources & documentation
Provider features were verified against official documentation. Accuracy figures are based on the test dataset described above.
Models & accuracy
- Radford et al. (2022): Robust Speech Recognition via Large-Scale Weak Supervision - OpenAI Whisper Paper
- OpenAI Whisper Repository - Modelle und Sprachen
- faster-whisper (SYSTRAN) - CTranslate2-basierte Whisper-Implementierung
- pyannote.audio 3.x - Speaker Diarization Pipeline
- Mozilla Common Voice - Offener Sprachdatensatz
Provider documentation
- Google Cloud Speech-to-Text - Funktionsübersicht und Spracherkennung
- Google Cloud Speech-to-Text - Speaker Diarization
- Microsoft Azure Speech Service - Übersicht und Dokumentation
- Azure Speech - Real-time diarization quickstart
- Amazon Transcribe - Entwicklerhandbuch
- Amazon Transcribe - Speaker partitioning (Diarization)
Ready for secure transcription?
Try scryp for free – encrypted from the very first second.