101–110 of 164 posts
Earlier quoted context omitted.
Wouldn't audible be perfectly positioned to take advantage of this. They have the perfect setup to integrate this into their offering.
It seems more likely that people will buy a digital copy of the book for a few bucks and then run the TTS themselves on devices they already own.
Is there something similar for STT? I’m using whisper distill models and they work ok. Sometimes it gets what I say completely wrong.
Parakeet is not really more accurate than Whisper, but it's much faster - faster than realtime even on CPU: https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3 . You have to use Nemo though, or mess around with third-party conversions. (Also has a big brother Canary: https://huggingface.co/nvidia/canary-1b-v2 . There's also the confusingly named/positioned Nemotron speech: https://huggingface.co/nvidia/nemotron-speech…
Of course, English only
Earlier quoted context omitted.
You can speak one language, switch to another language for one word, and continue speaking in the previous language.
But that's my point. You'll stop, switch, speak, stop, switch, resume. You're not going to be "I was in 東京 yesterday" as a single continuous sentence. It'll have to be broken up to three separate sentences spoken back to back, even for humans.
Good quality but unfortunately it is single language English only.
Agreed. I think they should have added the fact that it's English only in the title at the very least.
https://github.com/supertone-inc/supertonic
https://github.com/ekwek1/soprano
Haven't we had TTS for like 20+ years? Why does AI need to be shoved into it all of a sudden. Total waste of electricity.
[1] (2016 https://arxiv.org/abs/1609.03499) WaveNet: A Generative Model for Raw Audio
[2] (2017 https://arxiv.org/abs/1711.10433) Parallel WaveNet: Fast High-Fidelity Speech Synthesis
[3] (2021 https://arxiv.org/abs/2106.07889) UnivNet: A Neural Vocoder with Multi-Resolution Spectrogram Discriminators for High-Fidelity Waveform Generation
[4] (2022 https://arxiv.org/abs/2203.14941) Neural Vocoder is All You Need for Speech Super-resolution
The speed of improvement of tts models reminds me of early days of Stable Diffusion. Can't wait until I can generate audiobooks without infinite pain. If I was an investor I'd short Audible.
Set up SherpaTTS as the voice model for your phone (I like the en_GB-jenny_dioco-medium voice option, but there are several to choose from). Add a ebook to librera reader and open it. There's an icon with a little person wearing headphones, which lets you send the text continuously to your phone's tts, using just local processing on the phone. I don't have the latest phone but mine is able to process it faster than the audio is read, so the audio doesn't stop and start.
The voice isn't totally human sounding, but it's a lot better than the microsoft sam days, and once you get used to it the roboticness fades into the background and I can just listen to the story. You may get better results with kokoro (I couldn't get it running on my phone) or similar tts engines and a more powerful phone.
One thing I like about this setup is that if you want to swap back and forth between audio and text, you can. The reader scrolls automatically as it makes the audio, and you can pause it, read in silence for a while yourself and later set it going from a new point.
I just tried some sample verses, sounds natural.
But there seems to be a bug maybe? Just for fun, I had asked it to play the Real Slim Shady lyrics. It always seems to add 1 extra "please stand-up" in the chorus. Anyone see that?
It'd be great if it supported stdin&stdout for text and wav. Then it could get piped right into afplay
Earlier quoted context omitted.
You can do the same thing with Firefox' Reader Mode. On Linux you have to set up speech-dispatcher to use your favorite TTS as a backend.Once it is set up, there will be an option to listen the page.
Firefox should integrate that in their Reader Mode (the default System Voices are often very un-listable). Would seems like an easy win, and it's a non-AI feature so not polarising.
Speech-dispatcher commonly uses espeak-ng, which sounds robotic but is reportedly better for visually impaired users, because at higher speeds it is still intelligible. This allows visually impaired users to hear UI labels more quickly. For non visually impaired users, we generally want natural sounding voices and to use TTS in the same way we would listen to podcasts or a bedtime story.
With this system, users are in full control and can swap TTS models easily. If a model is shipped and, two weeks later, a smaller, newer, or better one appears, their work would become obsolete very quickly.
Earlier quoted context omitted.
An all-TTS audiobook offering is just about as appealing as an all-stable-diffusion picture gallery (that is, not at all).
Isn’t it more like an art gallery of prints of paintings? The primary art is the text of the book (like the painting in the gallery), TTS (and printing a copy) are just methods of making the art available.
To me, what you're saying is the same as saying the art of a movie is in the script, the video is just the method of making it available. And I don't think that's a valid take