Eleven Labs (EL) have just introduced a faster TTS model, Flash 2.5, which is specifically geared towards conversational agents.
See their announcement here –
https://www.youtube.com/watch?v=0YmHnkTVkFA
According to EL, speed is faster & quality is lower.
We reviewed the new model and can confirm that in the SitePal environment overall latency is reduced by 15% to 20% on average for English. For non-English improvement is much greater – with average latency reduced by 50% to %60.
As the new model is multilingual it can be used for all supported languages, which is why latency for non-English is more significantly improved. Until now, non-English input was processed using the slower multilingual model.
We were not able to audibly detect loss of quality. We concluded that the difference in quality is not meaningful in an online conversation scenario.
We have therefore modified the API to use the new Flash v2.5 model as its default model for EL, which means it will be used if you do not specify a different model when calling the API.
This update is now implemented. There is nothing you need to do if you use the default engine.
To review for yourself – check here – https://elevenlabs.io/app/speech-synthesis/text-to-speech (select model at top right)
If you prefer not to use the new default model, specify the model name in the xdata1 parameter when calling sayText or sayAI. To review this and other options for fine tuning EL audio generation, see details in the SitePal API reference. Check out the parameters for the sayText or sayAI functions & look for ‘xdata1’.
New model:
model_id=eleven_flash_v2_5
Previously used model – for English:
model_id=eleven_turbo_v2
Previously used model – for non-English:
model_id=eleven_multilingual_v2
Eleven Labs (EL) TTS can be integrated with SitePal by adding your EL API key to your SitePal ‘Connect’ page, and is one of several 3rd party TTS providers available for use with SitePal Avatars, to complement the built in TTS voices. Using 3rd party TTS requires the Platinum Plan.
