Speech synthesis: Difference between revisions

From OLPC
Jump to navigation Jump to search
Line 95: Line 95:


The OLPC XO platform, priced in the hundreds of dollars per device, already contains many of the base components needed, evident in the text-to-speech synthesis activity [[Speak]].
The OLPC XO platform, priced in the hundreds of dollars per device, already contains many of the base components needed, evident in the text-to-speech synthesis activity [[Speak]].
The OLPC icon-to-speech need has been expressed by many people independently, including discussions at [[Speak#Accessibility]] and [[Talk:Accessibility#Augmentative_and_Alternative_Communication]].
The OLPC icon-to-speech need has been expressed by many people independently, including discussions at [[Talk:Speak#Accessibility]] and [[Talk:Accessibility#Augmentative_and_Alternative_Communication]].


It appears that a proof of concept could be developed with a small time investment, and potential users are ready to test as soon as this is complete.
It appears that a proof of concept could be developed with a small time investment, and potential users are ready to test as soon as this is complete.

Revision as of 05:38, 19 March 2008

This article is a stub. You can help the OLPC project by expanding it.


Scope

This article is for collecting ideas and resources for using text-to-speech (TTS) speech synthesis on the XO.

Applications of Speech Synthesis wrt OLPC

Speech synthesis will not only be useful in improving the accessibility of the laptop but also for providing learning aides to the student.

Some simple educational activities that would benefit from the speech synthesis project include:

  • Pronounce - An activity teaching the child how to pronounce words correctly. It can be scaled up in the future to use speech recognition/ analysis of audio files to take audio input from the student. Based upon analysis and comparisons of the input audio file the activity can suggest appropriate corrections in the way the child speaks.
  • Story Book Reader - The Read activity can double up as an activity that would read stories that the child downloads on his/her XO. Children can be encouraged to read more and learn as much as they can. Learning through listening has its own advantages when compared to learning through reading and ad-hoc experimentation.
  • Listen and Spell - Students can listen to the XO speak a word. They must then spell the word and see if they did so correctly. This can be scaled up to a multiplayer game where students can challenge other students in their area. edit: Check out wiki.laptop.org/go/talkntype for beginning work in this area.
  • Talking Chatbots - Kids would love to shoot questions to an AI chatbot and hear it answer
  • Accessibility - Speech Synthesis tools are an integral component of software meant to improve accessibility. See http://live.gnome.org/Orca Orca] for more info.

Also see the following article which is a good read for the present context: Effective Adult Literacy Program

Existing software

Speak

Type text, and a funny face speaks what you typed. Pitch, speed, glasses, and mouth are adjustable.

Others

There are FOSS Free Open Source Software Speech-Synthesis packages which run on devices comparable to the XO. We are much more concerned with localization than is typical. And dialects can be a political issue. But TTS would help with Accessibility. And could be very cool.

Speech synthesis has a set of complex tradoffs of synthesizer size versus fidelity versus effort to localize a new language. The Wikipedia speech synthesis article discusses software that is available, which includes festival, flite, and espeak.

Espeak is small enough for us to often bundle and covers quite a few languages: ~10 languages currently supported tuned by native speakers. Localization to ten more languages is underway.

Synthesis is essential for accessibility to content by people with vision problems, and will need to be integrated with the ATK library used, as well as literacy training, other uses as part of a GUI. Full localization therefore involves selection of a suitable synthesis system and integration into the ATK framework, along with localization of that system for the particular language involved.

Speech synthesis is usually not a good guide for pronunciation – but it may be better than a poor teacher who has never had the opportunity to learn from a native speaker of that language.

eSpeak

eSpeak is currently included on the xo. .. But does not work directly to the sound card since the XO uses ALSA instead of OSS as its main Sound System,and enabling OSS Emulation in ALSA is not yet the default. Manually configuring your XO to emulate OSS in ALSA will provide the system devices that you require and allow full espeak functionality - Dking

If you are lacking OSS Emulation on your XO's sound sytstem setup in ALSA, some text can be played by piping espeak's standard output to another file:

$ espeak --stdout "Ello world." | gst-launch fdsrc fd=0 ! wavparse ! alsasink
$ espeak --stdout -vpt "Bem-vindo ao wiki da OLPC" | gst-launch fdsrc fd=0 ! wavparse ! alsasink
$ espeak --stdout "Using aplay." | aplay -

However, for some initial sounds, espeak fails to output valid audio to standard out (Trac #4002) . This includes letters c, h, k, p, q, t, v, z and possibly others. For example, this won't work: **this seems to work these days**

$ espeak --stdout "hello world." | aplay

A workaround is to first write the output to a file, then play back the file:

$ espeak -w temp.wav "hello world."; aplay temp.wav


Screen Reader is a DBus interface that allows the XO to use eSpeak via Python.

Festival

Flite is not currently included on the xo. Unless that changes, it would have to come out of your activity's space budget.

First, run /sbin/init 3 so yum doesn't run out of memory.  After yum, reboot.
$ yum install flite
$ flite -t 'Hello, world!'
Does it always sound this bad, or is just the default voice that works poorly? MitchellNCharity 16:42, 22 October 2007 (EDT)

Festival is not currently included on the xo. Unless that changes, it would have to come out of your activity's space budget.

First, run /sbin/init 3 so yum doesn't run out of memory.  After yum, reboot.
$ yum install festival
$ echo 'Hello, world!' | festival --tts

OLPCIconSpeak, under development

Overview

The goal of OLPCIconSpeak is to provide a low-cost assistive / augmentative communication tool for people with speech, motor, and/or developmental challenges. The immediate opportunity is to create open source software for the OLPC XO to allow a user to select concepts through a menu of icons, and synthesize speech from those selected concepts.

Existing tools in use for this purpose are expensive and proprietary. One example is DynaVox http://dynavoxtech.com/products/Series5_software.aspx , priced in the thousands of dollars per device.

The OLPC XO platform, priced in the hundreds of dollars per device, already contains many of the base components needed, evident in the text-to-speech synthesis activity Speak. The OLPC icon-to-speech need has been expressed by many people independently, including discussions at Talk:Speak#Accessibility and Talk:Accessibility#Augmentative_and_Alternative_Communication.

It appears that a proof of concept could be developed with a small time investment, and potential users are ready to test as soon as this is complete.

User Interface Design

Initial discussions suggest a user interface which allows users to navigate a hierarchy of basic concepts, allowing some variability of detail / zoom, due to the variability of users' motor skills used to select concepts.

3 levels of hierarchy at 7 +/-2 groups/concepts per level would allow selection among hundreds of concepts, which appears to be a useful balance between richness of expression and speed of selection.

Display and navigation of the hierarchy can be a combination of existing concentric & zoomable menu approaches:

We envision three such navigation areas, displayed from left to right across the screen, for the selection of a subject, a verb, and an object of a basic sentence, with no attempt at grammatical accuracy.

Conceptual Content

The concept hierarchy can be synthesized from a careful blend of existing taxonomies. For an initial proof of concept, two useful taxonomies are from sign language and the food pyramid. Use of sign language extends all the way to toddlers, as an increasingly popular supplemental communication before they develop speech abilities, such as the "Sign With Your Baby" materials. 100 basic signs provide some of the most useful concepts for basic living: http://www.lifeprint.com/asl101/pages-layout/concepts.htm . Sign language may be doubly useful in some cases, when motor skills allow for communication with the manual signs. Icon libraries are already established for American Sign Language, and readily available for many of the USDA food pyramid categories: http://openclipart.org/media/tags/vegetable .

Additional Enhancements and Uses

  • Input devices:
  • Additional languages & culturally-relevant icons
  • Add to & change the vocabulary & icons with photos, utilizing the built-in OLPC XO camera.
  • Run on smaller devices, such as mobile phones, music players, and PDAs with adequate speaker output.
  • Ability to operate with more grammatical correctness for more formal situations such as public and educational settings.
  • Teaching of reading & writing in native language.
  • Teaching of second or foreign languages.
  • Selectable foreign language or culture for speech output, enabling basic communication across languages or cultures.
  • Recording the selections as near-ontological content warrants further discussion.

User Interface mock-up, as a slide presentation

Open the slide presentation file: http://wiki.laptop.org/images/4/4d/IconSpeak_UI_text_demo_02.ppt .

[Work in progress: Icons are not drawn into this diagram yet. Will add actual icons soon, and storyboard this into a screencast. So for the moment, imagine that each word in black is replaced by an icon representing that concept.]

Click "people", "mom", "create", "cook", "food", and "beans", imagining the interface zooming in to where your pointer travels, for easier selectability.

Then the computer would consider your selections complete, and speak them.


Thanks for ideas contributed & discussed at PyCon 2008 by Tony Anderson, Lisa Beal, Annie Barkau, & Ed Cherlin.

- RMattB 2008 03 17

Please add your thoughts. :)

The state of the art

Commercial Text-To-Speech programs are getting very good now. The examples at the Digital Future Software Company site are very clear. They use AT&T technology and provide examples of Male and Female speech in English, French and Spanish. The XO needs open-source software that can approach this quality in a wide range of languages.--Ricardo 04:07, 17 August 2007 (EDT)

Resources

See also