Slashdot Mirror
Mozilla Releases Open Source Speech Recognition Model, Massive Voice Dataset (mozilla.org)
Mozilla's VP of Technology Strategy, Sean White, writes:
I'm excited to announce the initial release of Mozilla's open source speech recognition model that has an accuracy approaching what humans can perceive when listening to the same recordings... There are only a few commercial quality speech recognition services available, dominated by a small number of large companies. This reduces user choice and available features for startups, researchers or even larger companies that want to speech-enable their products and services. This is why we started DeepSpeech as an open source project.
Together with a community of likeminded developers, companies and researchers, we have applied sophisticated machine learning techniques and a variety of innovations to build a speech-to-text engine that has a word error rate of just 6.5% on LibriSpeech's test-clean dataset. vIn our initial release today, we have included pre-built packages for Python, NodeJS and a command-line binary that developers can use right away to experiment with speech recognition.
The announcement also touts the release of nearly 400,000 recordings -- downloadable by anyone -- as the first offering from Project Common Voice, "the world's second largest publicly available voice dataset." It launched in July "to make it easy for people to donate their voices to a publicly available database, and in doing so build a voice dataset that everyone can use to train new voice-enabled applications." And while they've started with English-language recordings, "we are working hard to ensure that Common Voice will support voice donations in multiple languages beginning in the first half of 2018."
"We at Mozilla believe technology should be open and accessible to all, and that includes voice... As the web expands beyond the 2D page, into the myriad ways where we connect to the Internet through new means like VR, AR, Speech, and languages, we'll continue our mission to ensure the Internet is a global public resource, open and accessible to all."
Together with a community of likeminded developers, companies and researchers, we have applied sophisticated machine learning techniques and a variety of innovations to build a speech-to-text engine that has a word error rate of just 6.5% on LibriSpeech's test-clean dataset. vIn our initial release today, we have included pre-built packages for Python, NodeJS and a command-line binary that developers can use right away to experiment with speech recognition.
The announcement also touts the release of nearly 400,000 recordings -- downloadable by anyone -- as the first offering from Project Common Voice, "the world's second largest publicly available voice dataset." It launched in July "to make it easy for people to donate their voices to a publicly available database, and in doing so build a voice dataset that everyone can use to train new voice-enabled applications." And while they've started with English-language recordings, "we are working hard to ensure that Common Voice will support voice donations in multiple languages beginning in the first half of 2018."
"We at Mozilla believe technology should be open and accessible to all, and that includes voice... As the web expands beyond the 2D page, into the myriad ways where we connect to the Internet through new means like VR, AR, Speech, and languages, we'll continue our mission to ensure the Internet is a global public resource, open and accessible to all."
Actually Web browsers need to implement a standardized speech recognition API (WebSpeech --- https://developer.mozilla.org/...), so this work could and probably will become part of Firefox. We wouldn't want speech-dependent Web applications to suck in Firefox on Linux because Firefox doesn't have access to a quality recognizer on free operating systems, would we?
This sort of thing is why building and maintaining Firefox is tremendously expensive. http://robert.ocallahan.org/20...
If - and I don't yet know if this is the case, they don't actually seem to say - this represents a stand-alone, does-not-go-to-the-LAN-or-WAN speech-to-text system... with an error rate of 6.5% on English speech as claimed... then it's way more important than Yet Another Web Browser.
This is precisely the kind of thing projects like Mycroft need to become not just another way to send your activity out on the net, which inherently decreases both reliability and security.
If indeed this is what this is, then the door opens for all manner of sophisticated home advances we can actually trust and depend on.
They claim around 1:1 [decode rate : normal speech rate] with a reasonably modern CPU/GPU. That needs considerable improvement. Reference quote from here:
That's a lot of computing power to hand off, particularly in a laptop. Using just the CPU, you'll be pegging it the whole time you're talking, and then some. For a decent desktop, it's at least doable, but it's still a very heavy compute load.
Though... saying "MacBook Pro" doesn't really tell us enough... I have a MacBook Pro that is a dual-core Intel machine... it's not what you'd call quick. There are a lot of different hardware configs that could be described by "MacBook Pro."
Seems like a pretty big deal to have to dedicate a server to the STT task (but then again, if I could get my STT tasks out from under the cloud... I'd probably do it. I have a spare 3 GHz 8-core hanging around, so...) but I think for general use, they have to do better. This isn't going to fly well on a Raspberry pi, for instance, it'll just get way behind.
Still. IMHO, this may be important. Very.
I've fallen off your lawn, and I can't get up.