AI voice synthesis has crossed a quality threshold in the last 18 months. Most listeners can't distinguish a top-tier AI voice from a human voice actor in a one-sentence test. In a five-minute listen, though, the tells show up. Here's what to listen for.
The five tells of bad dubbing
- Monotone prosody — the voice doesn't vary in pitch the way humans do when they're excited, sad, or making a joke.
- Wrong rhythm — sentences in some languages are longer than others, but the dub plays them in the same length as the original. A 12-word English sentence translated to Hindi becomes a 30-word Hindi sentence that's been compressed to fit.
- Misplaced stress — important words in a sentence are spoken too quietly or at the wrong pitch.
- Speaker bleed — when two characters are talking, the dub sometimes mixes them into a single voice.
- No breathing — real humans take breaths at natural places. Cheap text-to-speech engines skip them, which sounds uncanny.
How SyncDub compares
We use a single voice across all 70+ languages. That's a deliberate trade-off: it means our voice is consistent (you won't get whiplash from one language to the next), but it also means we can't replicate every regional accent perfectly. We've optimised for natural prosody and natural breath placement — the two tells that listeners notice most.
How to evaluate quality yourself
- Listen to a 60-second clip with your eyes closed. If you can tell it's AI within 5 seconds, the voice isn't good enough.
- Translate a song. If the dub ignores the melody entirely, the model is using the wrong TTS for the task.
- Translate a heated argument. If both speakers sound the same, the model isn't doing speaker diarisation.
- Translate a joke. If the punchline lands at the wrong beat, the rhythm is off.