Why is what I'm saying not being transcribed correctly?

Transcribing non-native speakers isn't easy. Depending on your pronunciation, you may occasionally see that it doesn't quite return what you intended. Frustrating at times.

Transcription models can also struggle with background noise, as well as long silences. To maximise accuracy, we recommend speaking in a quiet environment, and pausing the recording if you are not ready to speak (if you have 'auto-record' on).

If you are in call mode in a noisy environment, you may want to tap the mute icon when you're not speaking. Currently, transcription is more accurate in chat modes than call mode, as it's harder for AI to transcribe your speech in real-time (as required in call mode).

If you're having trouble being understood, consider hitting pause before sending your reply so you can check it before sending, and also ensure 'auto-send' is off.

For standard chat modes, the T1 transcription model is recommended as it:

• Is more accurate than any other model (across the languages we've tested)
• Has an impressive ability to understand even if you switch languages half way through a sentence
• Is optimised by our team for speed
• Does not hallucinate
• Is quite good at filtering out background noise

The only real downside we've observed is that it will transcribe literally anything you say, which isn't ideal when you hesitate and say 'ummm...' or 'err..'. We do have a setting - 'Improve transcription accuracy' available in chat modes, which triggers an AI that aims to clean up such hesitations.

All providers occasionally have technical issues. We built an automatic retry system so that if one model fails, your recording will be sent to another model to retry. It's not bulletproof, but it helps.

Our team experiments with new models whenever they come out, so that we can continue to offer the best available tech.

We're also looking at fine-tuning models for languages that are less widely spoken, and therefore have less training data and lower accuracy.