I’m a firefighter who’s also a software engineer and am working on a training app. In it, I have it generating text dispatches of various scenes for us to discuss on rainy slow training days. Such as “Respond to 123 Main st, for a report of a smell of smoke” etc.
I already use google maps to generate a random address and show the map / street view. With the maps api I can domain lock the key…
But with their Text to Speech api I cannot. Seems silly. But I get it.
Are there any alternatives? I would be ok spinning up a middle server to also rework the audio to generate radio static, etc, but first pass I am looking for non-robotic (ie not browser based) TTS.
Thoughts?
- eSpeak is the old favorite. its very robotic, but with a lot of fiddling, you can get the option that will sound like a person.
- piper https://noerguerra.com/how-to-read-text-aloud-with-piper-and-python/ is pretty good at mimicking human speach. Still sorta robotic, but can run on just about anything.
- https://github.com/suno-ai/bark is hard to work with, but will definitely get you a human sounding voice if that is what you are looking for.
Sometime last year, I was trying to get an open source TTS to generate audiobooks from books that will most likely never get a real person attached. I ended up using piper for it. It was “good enough” for car rides and such. Bark felt kinda like cheating/uncanny valley and would do strange things to the voice after a while.
Hope that helps!
if i read this correctly a piper has webasm bindings too? I am running the training app in a browser as a vue3+ts stack. So I could spin up a worker on cloudflare, api key lock, and then use that or run it all browser side with wasm.
Time to tinker!
Give it a shot. Its been a full year since ive messed with this so let me know how it goes. Might be some better stuff out now. But yeah piper/eSpeak both have fantastic performance. You can even put it on your phone should the need arise: https://f-droid.org/packages/org.woheller69.ttsengine/
cool! i know cf workers ai can run a tts model and i can domain lock the api key… but this seems cooler. :)
sweet. thanks!
I’ve really enjoyed using Kokoro for generating audiobooks:
Be sure to first try using this convenient API wrapper:
Note that not all the modelled voices in Kokoro-82M are of equal quality, given disparities in limited training data from reference speakers. However, what’s cool is that you can prescribe polynomial weights to multiple voices tags, enabling you to synthesize different variants weighted more heavily from the highest quality voices.
- https://huggingface.co/hexgrad/Kokoro-82M
- https://huggingface.co/hexgrad/Kokoro-82M/blob/main/VOICES.md#american-english
One current limitation for Kokoro is that there’s no way to prescribe emotion or intonation procedurally using markup tags like SSML in the source text, unlike other models like Orpheus. But Orpheus sometimes generate weird hallucinations like repeating sentences, injecting new phrases, appending radio silence or filter words, and generally increasing the tempo of words per minute as a sentence progresses. Still, this may be of interest if you want to add emotion like fear or urgency to your generated dispatches, and manage to tune the input temperature you want for the model.
However, Kokoro is a lot more compute efficient and audibly consistent, requiring less scrutiny or manual supervision. The author behind Kokoro now also looks to be working towards an emotional variant as well:
Reference project I’ve been following for audiobook generation:
wow. thanks!
