Report Bug / Feature Request

AI Models

Choose Your Transcription Engine — Compare accuracy, speed, and language support across leading speech recognition models.

STT.ai Enhanced
STT.ai
Recommended STT.ai

STT.ai's flagship speech-to-text model with best-in-class accuracy and speed. Optimized for production workloads.

3.2%
WER
100
Langs
160.0x
Speed
Proprietary Try it →
Whisper Large V3
OpenAI
Recommended

OpenAI's largest and most accurate Whisper model. Excellent multilingual support with 99 languages.

4.2%
WER
99
Langs
8.0x
Speed
MIT Try it →
Whisper Turbo
OpenAI
Recommended

OpenAI's speed-optimized Whisper variant. 4x faster than Large V3 with minimal accuracy loss.

5.1%
WER
99
Langs
32.0x
Speed
MIT Try it →
NVIDIA Canary
NVIDIA
Recommended

NVIDIA's multi-task ASR model with top-tier accuracy on English. Built on the NeMo framework.

3.5%
WER
4
Langs
45.0x
Speed
CC-BY-4.0 Try it →
Moonshine
Useful Sensors

Ultra-lightweight ASR model designed for edge devices. Runs on Raspberry Pi with minimal latency.

7.8%
WER
1
Langs
80.0x
Speed
MIT Try it →
NVIDIA Parakeet
NVIDIA
Recommended

NVIDIA's CTC-based English ASR model. One of the most accurate open-source English models available.

3.0%
WER
1
Langs
55.0x
Speed
CC-BY-4.0 Try it →
SenseVoice
FunAudioLLM

Multilingual speech understanding model with emotion recognition and audio event detection.

5.5%
WER
50
Langs
50.0x
Speed
MIT Try it →
Distil-Whisper
Hugging Face

Distilled version of Whisper Large V3. 6x faster with 49% fewer parameters while maintaining accuracy.

5.8%
WER
99
Langs
48.0x
Speed
MIT Try it →
Vosk
Alpha Cephei

Lightweight offline speech recognition. Works without internet, ideal for privacy-sensitive and embedded use cases.

12.0%
WER
20
Langs
100.0x
Speed
Apache 2.0 Try it →

Hvordan velge den riktige modellen

Forskjellige utskriftsmodeller utmerker seg på forskjellige områder. Bruk denne veiledningen for å velge den modellen du vil bruke.

Model WER Speed Språk Beste for
STT.ai Enhanced 3.2% 160.0x 100 STT.ai's flagship speech-to-text model with best-in-class accuracy and speed. Optimized …
Whisper Large V3 4.2% 8.0x 99 OpenAI's largest and most accurate Whisper model. Excellent multilingual support …
Whisper Turbo 5.1% 32.0x 99 OpenAI's speed-optimized Whisper variant. 4x faster than Large V3 with …
NVIDIA Canary 3.5% 45.0x 4 NVIDIA's multi-task ASR model with top-tier accuracy on English. Built …
Moonshine 7.8% 80.0x 1 Ultra-lightweight ASR model designed for edge devices. Runs on Raspberry …
NVIDIA Parakeet 3.0% 55.0x 1 NVIDIA's CTC-based English ASR model. One of the most accurate …
SenseVoice 5.5% 50.0x 50 Multilingual speech understanding model with emotion recognition and audio event …
Distil-Whisper 5.8% 48.0x 99 Distilled version of Whisper Large V3. 6x faster with 49% …
Vosk 12.0% 100.0x 20 Lightweight offline speech recognition. Works without internet, ideal for privacy-sensitive …

Hva er WER (ordfeilrate)?

Ordfeilrate (WER) er standardmål for måling av talegjenkjenningsnøyaktighet. Det beregner prosentvis antall ord i en utskrift som er forskjellig fra referansen. En WER på 5% betyr omtrent 5 av hver 100 ord inneholder en feil. Lavere er bedre.

Profesjonelle transkriptionister oppnår typisk en WER på 4-5%. De beste AI- modellene samsvarer nå eller nærmer seg nøyaktighet på menneskenivå på ren lyd.

Er du usikker på hvilken modell som skal brukes?

Forsøk vår standard – Whisper Large V3 Turbo gir den beste likevekten mellom fart og nøyaktighet. Fri til å starte, ikke nødvendig å logge inn.

Begynn å tranulere gratis