Why is generated speech sometimes glitchy?

RadioNewsAI uses ElevenLabs' voice generation technology to produce natural-sounding AI voices for your newscasts. It's widely considered to be the most reliable service of its kind available today; however, as AI is still in its infancy, it can occasionally produce unintended results.

With RadioNewsAI, we've tried to mitigate this as much as possible, such as allowing pronunciations to be set up and providing example scripts that the AI will train itself on to replicate your scripting preferences.

On this page, we'll explain some common issues you may encounter with AI-generated voices and how they can be rectified.

Inconsistency

Sometimes, the same script read out by an AI voice can sound different each time it is generated, even when using the same settings.

If you are using a custom voice which you have cloned from recordings, make sure you are using the highest-quality samples and have provided enough example audio for it to refer to. Five minutes of quality audio as reference material gives the AI more to work with than 30 seconds from an off-air recording.

Changing languages or accents

AI voices can sometimes switch languages or drift between accents, especially when reading longer stories. It is a known issue that ElevenLabs is currently working to resolve.

Glitchy or corrupted speech

There have been reports of AI voices occasionally repeating parts of sentences or producing muffled or garbled speech. ElevenLabs says "this is a very rare issue" but unfortunately, it's only suggestion is to re-generate any affected sections of audio.

To do this in RadioNewsAI, you will need to make a manual change to any stories which were distorted. This may just be re-phrasing a few words or changing punctuation.

This is required as RadioNewsAI re-uses existing generated audio where possible to avoid eating into your available seconds of generated voices. Unless the story content is changed, it will try to reuse any glitchy audio previously generated.