Initial release of GrapheneOS Speech Services for text-to-speech

mar2112

Used it for the first time today, with CoMaps, and received some interesting instructions. Turn onto us 101. "Us" not "you ess". And also turn onto Fairfax hyphen minus Bolinas road. Proper name of the street is Fairfax-Bolinas Rd.

Also had interesting pronunciations for Spanish street names. Madera was pronounced "Mad-er-a" instead of "Ma-der-a".

It to us a moment to understand what was being said the first time.... after that out just have is chuckles.

de0u

DohnJoe I was just thinking that for the time being, people could start recording their voices, to have the datasets set properly.

I think part of the problem is that people who have not yet built a text-to-speech system can imagine what builders would want as input data, but that may be very different from the needs of people who have built a text-to-speech system.

So while it might be useful for people to volunteer if they have prior experience developing text-to-speech systems, or have previously coordinated data collection for TTS developers, I suspect it's not currently productive for people to volunteer to record their voices in various languages.

Something that may be useful is for people to install the current GrapheneOS English TTS, to use it for various things, and to report issues ( as userofgos mar2112 have done). For example, does the voice sound good via headphones or via a car speaker system?

b_billy

mar2112 I used to use Petal Maps when I was still using CalyxOS, and always got a laugh out of, "At the traffic light, turn left on Vanderbilt Pakwee." (Vanderbilt Pkwy)

soupslurpr

mar2112 Thanks for the feedback! The "us 101" issue has been resolved and will be fixed in the next release.

The hyphen-minus behavior happens because "Fairfax-bolinas" isn't in the dictionary. This activates the fallback behavior, where a neural network is used to predict the pronunciation of the words, and other characters like hyphens are read out loud.

For the "Ma-der-a" behavior, that word isn't in the dictionary, so the neural network tries to guess its pronunciation. The model can sometimes get things wrong—especially with Spanish street names—because it was trained on the English dictionary the app uses.

DohnJoe

de0u Things like that might happen down the road, but I would be surprised if it happened in under a year.

Yea, I thought somewhat the same, I was just thinking that for the time being, people could start recording their voices, to have the datasets set properly.
I guess having some time to build momentum could help, to make sure contributors learn how to submit correct quality samples on the correct platforms.

I would say this might benefit not only GOS.
Being open data it would allow other projects to grow as well, so having some directions would be beneficial, I think.

If I had to record myself (you probably don't want my voice on your smartphone haha) I wouldn't know where to start, as in equipment and the rest.
I don't even know if there is some kind of test you can take to understand if your voice might be suitable for the task, before starting.

Johnnyloans

DohnJoe people could start recording their voices, to have the datasets set properly.

Common voice had people read random sentences from Wikipedia.

People should also be prepared to have all of their work go unused. Or upload it elsewhere.

you probably don't want my voice on your smartphone haha)
some kind of test you can take to understand if your voice might be suitable for the task

I'd imagine as long as you don't have an undesired accent.

I'd love to have your voice in my phone! Unless you have a cockney accent... Then maybe.

Audio training would alter the pitch, amplitude.
or add noise (for Speech to Text).
So, everyone's voice would sound bad to the "student" being trained.

Johnnyloans

Johnnyloans People should also be prepared to have all of their work go unused.

There is a chance of this happening somewhere between 0%-100%. I didn't mean to say it is guaranteed.

internetman

internetman Hehe lol to clarify I didnt mean to use the voices, more the other way around.

Pocketstar

This is really amazing! I use TTS and VTT (Voice To Text) alot, I use it along with a local LLM and some voice notifications. My girlfriend has special needs (blindness) so she will also be very happy to hear this. Right now she is working with a more bloated TTS and voice recognition, even though the underlying Talkback interface always works well for her. (She knows English as good as her native language)
Sadly she is less understanding of the importance of apps with a light footprint, but I'll try to explain it is beneficial for the speed, freedom and often privacy of software and battery life.
The devs are putting so much work into the fine details, sometimes I wonder how they can still have any free time left? But if it is their life's work, I can imagine it is their pride and joy, and the results really show and is extremely constructive! Furthermore, it is super important for whistleblowers, they help alot of people being safe.

Johnnyloans

Pocketstar Right now she is working with a more bloated TTS and voice recognition
Sadly she is less understanding of the importance of apps with a light footprint

Some things require the most expensive or strongest etc.

Phones hold 128GB, 256GB, 512GB, etc.
There isn't much of a difference between 300MB and 160MB for a very important tool.

Especially for a blind person, having the most accurate tool is best to avoid fixing mistakes.
It's similar for your LLMs. Running one that fits on 1GB of memory isn't super great.

With that said, this TTS sounds real nice.
When I am consulting someone, at the end of the day, they are the ones having to deal with the end-result so I leave it up to them.

Pocketstar

de0u To be clear, I don't believe the project's stated plan to support German and/or French first is based on any dislike of Dutch or any intention to not support Dutch.

I know I speak for many when I say Dutch people usually have English as their secondary language. We grew up with English TV shows that have Dutch subtitles, (which can serve as English education) and use alot of English games and software, speaking English in the Netherlands is as normal as speaking Dutch. They usually speak US-English rather that with a British dialect due to all the media being from the US.
I am Dutch myself, but I prefer English due to software compatibillity, so I usually set TTS en VTT to English regardless, most of my notifications are English as well.
Only some very old Dutch people don't know how to speak English, or speak what we call "Charcoal-English/steenkolen-engels." but I'm pretty sure 99% of those people do not have GrapheneOS. (Usually a Nokia they are familliar to or something like it)
My LLM's are terrible in Dutch, so I use those in English as well.

Panda-na

Johnnyloans
GrapheneOS have previously mentioned that it is important especially in the startup wizard when you first boot the phone. In that case, there is a limited set of predefined words/sentences in the wizard. There is no need to have the whole language speech model bundled with the OS. It can be an option to download after the wizard. In that way, there is only need for a few MB of audiofiles to be bundled. Then you can add several languages with basically no data impact.

Pocketstar

Johnnyloans Especially for a blind person, having the most accurate tool is best to avoid fixing mistakes.
It's similar for your LLMs. Running one that fits on 1GB of memory isn't super great.

True, and besides; she is perfectly capable of choosing herself what she finds the most fitting and convienient for herself, after all; she is the one having to use the tools to be able to identify the objects and text on the screen.
I'll tell her she can install the GrapheneOS Speech Services on her phone, and she can always check if she likes it.
I'm very happy with it myself though. (-:

Johnnyloans

Panda-na GrapheneOS have previously mentioned that it is important especially in the startup wizard when you first boot the phone. In that case, there is a limited set of predefined words/sentences in the wizard.

I'd have to see the context and what the message was. I doubt GOS would recommend for users to install small apps.

For the speech app, it makes sense to not have languages that many people won't use. Just have what's needed.

Panda-na There is no need to have the whole language speech model bundled with the OS.

Right, don't have to have several languages, but that is a different situation than:
I can choose 2 English only speech apps, one is 300MB and the other is 160MB. Personally, I would weigh other metrics other than which is smallest and that's the best.

Also, I think GOS speech app is the best sounding. There's a few quirks with version 1.0 saying US101 like someone else mentioned but that's fixable.

persimmonpomelo

Panda-na In that case, there is a limited set of predefined words/sentences in the wizard.

The speech services are for a screen reader like TalkBack. A predefined set of words/sentences would not be viable in that case. The set would have to include the TalkBack tutorial, all verbosity hints, etc. What it reads out can be arbitrary with user input like from the keyboard in the setup wizard.

Stock Pixel OS allows activation of Android Accessibility Suite TalkBack in the setup wizard, and Google speech services are pre-installed to provide a jumping off point to complete the entire setup process with a functional screen reader without the need for external assistance.

Developer-Dude

b_billy even Google map's voice navigation isn't perfect! It keeps saying Grande Prairie "Gron - day prairie" the proper way to say it is just "grand prairie". Its always a laugh. Especially since even Google can't get it! :)

MEGAbyter

Can I use it to set reminders for myself? Like ideally I'd like to be able to hold the power key and say "remind me in xxx to xxx" and have it send me a notification at the designated time..I used that all the time on my 24U and is the one thing I miss the most with GOS. Is there a way to accomplish that now? Or something similar?

Johnnyloans

MEGAbyter

No unfortunately.
This app is for having written word spoken out loud.

For your needs, it'd require the opposite -- speech changed to written word.
and then another app to receive the command and send it to your calendar or clock.

If you don't say the exact, magic phrase then it'd require more code for synonyms or an LLM AI to read what you said. (e.g. "remind me in... " or "tell me in xxx..." )

This is open source and should be able to be set as your assistant app

https://github.com/DicioTeam/dicio-android

I've never used it and perhaps there are others.

Pocketstar

For spoken notifications, people can install one of the following additional software:
https://github.com/pilot51/voicenotify
https://github.com/mitchib1440/SpeakThat
For local and remote LLM's, I use ChatterUI, which has a TTS setting build in. https://github.com/Vali-98/ChatterUI

Developer-Dude

Pocketstar i use Dicio as an open source assistant, but are there any better open source assistants? The only thing I like an assistant for is driving. Its easier (and safer) to press the voice button on my radio and say "call mom" or "navigate to xyz" than to pick up my phone and use it. Dicio is okay at this, bit not nearly great.

becomeparmesan

Developer-Dude You might be more lucky lucking for "agents" as that seems to be the newest terminology for that. I could find in F-Droid "Hermes Agent" and "Kai 9000", maybe then can do what you need.

Pocketstar

Developer-Dude I am unfamilliar with Dicio myself, but it looks very promising, I will look into it. I may use it myself if it comes in handy for my usecase.
However, I am not familliar with any other assistants, but perhaps the suggestions from becomeparmesan may indeed be a good idea to try out. (Hermes Agent and Kai 9000) I will give those a try as well. Just be cautious about prompt injection attacks like from websites with suggestions/prompts aimed at AI-agents. (Not sure if those attacks can be effective on GrapheneOS, or with the aformentioned agents, but I doubt it can do any serious harm though...)

« Previous Page Next Page »