So everyone's doing their own thing after the 'there was an attempt' at HTML5 spec definition??
Mozilla Overhauls Speech-To-Text Contribution Interface
11–20 of 46 posts
Re: Mozilla Overhauls Speech-To-Text Contribution Interface
#12Earlier quoted context omitted.
That's what the second pass is for, right, to screen out actually unintelligible or misrecorded entries. The English (in)fluency is more of a feature, though, than a bug. The goal isn't to produce a speech-to-text system that can recognize a perfectly miked BBC announcer. It's to be able to recognize a wide variety of people speaking fairly naturally in imperfect conditions, using whatever accent they use for casual…
> "The goal isn't to produce a speech-to-text system that can recognize a perfectly miked BBC announcer." Wait what? The headline is about text-to-speech aka speech synthesis, not speech recognition (speech-to-text.) Are they trying to do both? It seems to me that you'd train both using different sorts of datasets. If you wanted TTS to be intelligible to the most number of people, training to to speak like a 'perfect…
The headline was inaccurate (now fixed) — the Mozilla Voice project is about speech recognition aka STT not TTS.
It would be kinda interesting to have a TTS system learn from a neighboring STT system so that it gradually adopts your accent, though. I'm not sure if that would be more usable but it would be an interesting experience.
Re: Mozilla Overhauls Speech-To-Text Contribution Interface
#13Did they consider to use MTurk for this?
If you think about it, this is a Mechanical Turk approach, just done for free with volunteers not with the financial impetus of MTurk. That's why they've given particular attention to the second pass over the data.
If you want to read more about it, the GitHub repo and in particular the issues cover a lot of the obvious questions like this. They're here: https://github.com/mozilla/voice-web/issues
Re: Mozilla Overhauls Speech-To-Text Contribution Interface
#14Earlier quoted context omitted.
That's what the second pass is for, right, to screen out actually unintelligible or misrecorded entries. The English (in)fluency is more of a feature, though, than a bug. The goal isn't to produce a speech-to-text system that can recognize a perfectly miked BBC announcer. It's to be able to recognize a wide variety of people speaking fairly naturally in imperfect conditions, using whatever accent they use for casual…
> "The goal isn't to produce a speech-to-text system that can recognize a perfectly miked BBC announcer." Wait what? The headline is about text-to-speech aka speech synthesis, not speech recognition (speech-to-text.) Are they trying to do both? It seems to me that you'd train both using different sorts of datasets. If you wanted TTS to be intelligible to the most number of people, training to to speak like a 'perfect…
You were quite right that in the way Mozilla are using this, it's suited to speech to text, but running it locally is actually a reasonable way to record a single speaker set of data for use in TTS training.
That's exactly what I've been doing recently, and using it with https://github.com/r9y9/deepvoice3_pytorch/blob/master/READM... is providing reasonably good results - it definitely has my intonation (if somewhat crossed with a Dalek!!)
Re: Mozilla Overhauls Speech-To-Text Contribution Interface
#15My biggest issue with this project is that more than half of the contributions are from people who have failed to record correctly, or who are not fluent in English.
I'm more interested in seeing how a bucket as big as 'english' turns out when you put all accents from around the world together. It would be another interesting project to separate them. I home they're at least saving the browser locale or GeoIP info to help categorize it.
Re: Mozilla Overhauls Speech-To-Text Contribution Interface
#16My biggest issue with this project is that more than half of the contributions are from people who have failed to record correctly, or who are not fluent in English.
I find the contributions where the person doesn’t speak the recording at all or appears to be intentionally not reading the corpus to be more frustrating than those where the speaker is not perfectly fluent.
Re: Mozilla Overhauls Speech-To-Text Contribution Interface
#17So everyone's doing their own thing after the 'there was an attempt' at HTML5 spec definition??
What?
Re: Mozilla Overhauls Speech-To-Text Contribution Interface
#18Am using Brave on Android (basically chrome with afblocker). I accidentally mislabeled some (probably) correct samples because the waves move when you click the play button, even if the audio hasn't loaded yet and I thought it was just a blank recording.
Re: Mozilla Overhauls Speech-To-Text Contribution Interface
#19From the description:
Common Voice is a project to help make voice recognition open to everyone. Now you can donate your voice to help us build an open-source voice database that anyone can use to make innovative apps for devices and the web.
I'll be the first to note that here's another piece of personally identifying information you just "donated"...
Re: Mozilla Overhauls Speech-To-Text Contribution Interface
#20This is a nicely designed interface. Well done, Mozilla. Validating sentences is quite fun, listening to different accents from around the world. Try it out if you haven't already: https://voice.mozilla.org/en/listen It's awesome that the dataset is offered with a CC-0 license: https://voice.mozilla.org/en/data , does anyone know if it includes the answers from the survey? I have a limited bandwidth internet, so I ha…
Trolling of crowdsourced data isn't unheard of.
...in that case it was specifically done in retaliation to Google trying to get free mental labour from ReCAPTCHA users in return for being able to post to 4chan. Quite a different situation.