- From: Martin Duerst <duerst@w3.org>
- Date: Fri, 31 Jan 2003 22:49:40 -0500
- To: www-voice@w3.org
- Cc: w3c-i18n-ig@w3.org
Dear Voice Browser WG,
These are the Last Call comments on Speech Synthesis
Markup Language (http://www.w3.org/TR/speech-synthesis/)
from the Core Task Force of the Internationalization (I18N) WG.
Please make sure that you send all emails regarding these
comments to w3c-i18n-ig@w3.org, rather than to me personally
or just to www-voice@w3.org (to which we are not subscribed).
These comments are based on review by Richard Ishida and myself and
have been discussed and approved the last I18N Core TF teleconference.
They are ordered by section and numbered for easy reference.
We have not classified these issues into editorial and substantial,
but we think that it should be clear from their discription.
General:
[01] For some languages, text-to-speech conversion is more difficult
than for others. In particular, Arabic and Hebrew are usually
written with none or only a few vowels indicated. Japanese
often needs separate indications for pronunciation.
It was no clear to us whether such cases were considered,
and if they had been considered, what the appropriate
solution was.
SSML should be clear about how it is expected to handle these
cases, and give examples. Potential solutions we came up with:
a) require/recommend that text in SSML is written in an
easily 'speakable' form (i.e. vowelized for Arabic/Hebrew,
or with Kana (phonetic alphabet(s)) for Japanese. (Problem:
displaying the text visually would not be satisfactory in this
case); b) using <sub>; c) using <phoneme> (Problem: only
having IPA available would be too tedious on authors);
d) reusing some otherwise defined markup for this purpose
(e.g. <ruby> from http://www.w3.org/TR/ruby/ for Japanese);
e) creating some additional markup in SSML.
General: Tagging for bidirectional rendering is not needed
[02] for text-to-speech conversion. But there is some provision
for SSML content to be displayed visually (to cover WAI
needs). This will not work without adequate support of bidi
needs, with appropriate markup and/or hooks for styling.
General: Is there a tag that allows to change the language in
[03] the middle of a sentence (such as <html:span>)? If not,
why not? This functionality needs to be provided.
Abstract: 'is part of this set of new markup specifications': Which set?
[04]
Intro: 'The W3C Standard' -> 'This W3C Specification'
[05]
Intro: Please shortly describe the intended uses of SSML here,
[06] rather than having the reader wait for Section 4.
Section 1, para 2: Please shortly describe how SSML and Sable are
[07] related or different.
1.1, table: 'formatted text' -> 'marked-up text'
[08]
1.1, last bullet: add a comma before 'and' to make
[09] the sentence more readable
1.2, bullet 4, para 1: It might be nice to contrast the 45 phonemes
[10] in English with some other language. This is just one case that
shows that there are many opportunities for more internationally
varied examples. Please take any such oppurtunities.
1.2, bullet 4, para 3: "pronunciation dictionary" ->
[11] "language-specific pronunciation dictionary"
1.2: How is "Tlalpachicatl" pronounced? Other examples may be
[12] St.John-Smyth (sinjen-smaithe) or Caius College
(keys college), or President Tito (sutto) [president of the
republic of Kiribati (kiribass)
1.1 and 1.5: Having a 'vocabulary' table in 1.1 and then a
[13] terminology section is somewhat confusing.
Make 1.1 e.g. more text-only, with a reference to 1.5,
and have all terms listed in 1.5.
1.5: The definition of anyURI in XML Schema is considerably wider
[14] than RFC 2396/2732, in that anyURI allows non-ASCII characters.
For internationalization, this is very important. The text
must be changed to not give the wrong impression.
1.5 (and 2.1.2): This (in particular 'following the
[15] XML specification') gives the wrong impression of where/how
xml:lang is defined. xml:lang is *defined* in the XML spec,
and *used* in SSML. Descriptions such as 'a language code is
required by RFC 3066' are confusing. What kind of language code?
Also, XML may be updated in the future to a new version of RFC
3066, SSML should not restrict itself to RFC 3066
(similar to the recent update from RFC 1766 to RFC 3066).
Please check the latest text in the XML errata for this.
2., intro: xml:lang is an attribute, not an element.
[16]
2.1.1, para 1: Given the importance of knowing the language for
[17] speech synthesis, the xml:lang should be mandatory on the root
speak element. If not, there should be a strong injunction to use it.
2.1.1: 'The version number for this specification is 1.0.': please
[18] say that this is what has to go into the value of the 'version'
attribute.
2.1.2., for the first paragraph, reword: 'To indicate the natural
[19] language of an element and its attributes and subelements,
SSML uses xml:lang as defined in XML 1.0.'
The following elements also should allow xml:lang:
[20] - <prosody> (language change may coincide with prosody change)
- <audio> (audio may be used for foreign-language pieces)
- <desc> (textual description may be different from audio,
e.g. <desc xml:lang='en'>Song in Japanese</desc>
- <say-as> (specific construct may be in different language)
- <sub>
- <phoneme>
2.1.2: 'text normalization' (also in 2.1.6): What does this mean?
[21] It needs to be clearly specified/explained, otherwise there may
be confusion with things such as NFC (see Character Model).
2.1.2, example 1: Overall, it may be better to use utf-8 rather than
[22] iso-8859-1 for the specification and the examples.
2.1.2, example 1: To make the example more realistic, in the paragraph
[23] that uses lang="ja" you should have Japanese text - not an English
transcription, which may not use as such on a Japanese text-to-speech
processor. In order to make sure the example can be viewed even
in situations where there are no Japanese fonts available, and
can be understood by everybody, some explanatory text can provide
the romanized from. (we can help with Japanese if necessary)
2.1.2, 1st para after 1st example: Editorial. We prefer "In the
[24] case that a document requires speech output in a language not
supported by the processor, the speech processor largely determines
the behavior."
2.1.2, 2nd para after 1st example: "There may be variation..."
[25] Is the 'may' a keyword as in rfc2119? Ie. Are you allowing
conformant processors to vary in the implementation of xml:lang?
If yes, what variations exactly would be allowed?
2.1.3: 'A paragraph element represents the paragraph structure'
[26] -> 'A paragraph element represents a paragraph'. (same for sentence)
Please decide to either use <p> or <paragraph>, but not both
(and same for sentence).
2.1.4: <say-as>: For interoperability, defining attributes
[27] and giving (convincingly useful) values for these attributes
but saying that these will be specified in a separate document
is very dangerous. Either remove all the details (and then
maybe also the <say-as> element itself), or say that the
values given here are defined here, but that future versions
of this spec or separate specs may extend the list of values.
[Please note that this is only about the attribute values,
not the actual behavior, which is highly language-dependent
and probably does not need to be specified in every detail.]
2.1.4, interpret-as and format, 6th paragraph: requirement that
[28] text processor has to render text in addition to the indicated
content type is a recipe for bugwards compatibility (which
should be avoided).
2.1.4, 'locale': change to 'language'.
[29]
2.1.4: How is format='telephone' spoken?
[30]
2.1.4: Why are there 'ordinal' and 'cardinal' values for both
[31] interpret-as and format?
2.1.4 'The detail attribute can be used for all say-as content types.'
[32] What's a content type in this context?
2.1.4 detail 'strict': 'speak letters with all detail': As opposed
[33] to what (e.g. in that specific example)?
2.1.4, last table: There seem to be some fixed-width aspects in the
[34] styling of this table. This should be corrected to allow complete
viewing and printing at various overall widths.
2.1.4, 4th para (and several similar in other sections):
[35] "The say-as element can only contain text." would be easier
to understand; we had to look around to find out whether the
current phrasing described an EMPTY element or not.
2.1.4. For many languages, there is a need for additional information.
[36] For example, in German, ordinal numbers are denoted with a number
followed by a period (e.g. '5.'). They are read depending on case
and gender of the relevant noun (as well as depending on the use
of definite or indefinite article).
2.1.4, 4th row of 2nd table: I've seen some weird phone formats, but
[37] nothing quite like this! Maybe a more normal example would NOT
pronounce the separators. (Except in the Japanese case, where the
spaces are (sometimes) pronounced (as 'no').)
2.1.5, <phoneme>:
[38] It is unclear to what extent this element is designed for
strictly phonemic and phonetic notations, or also (potentially)
for notations that are more phonetic-oriented than usual writing
(e.g. Japanese kana-only, Arabic/Hebrew with full vowels,...)
and where the boundaries are to other elements such as <say-as>
and <sub>. This needs to be clarified.
2.1.5 There may be different flavors and variants of IPA (see e.g.
[39] references in ISO 10646). Please make sure it is clear which
one is used.
2.1.5 IPA is used both for phonetic and phonemic notations. Please
[40] clarify which one is to be used.
2.1.5 This may need a note that not all characters used in IPA are
[41] in the IPA block.
2.1.5 This seems to say that the only (currently) allowed value for
[42] alphabet is 'ipa'. If this is the case, this needs to be said
very clearly (and it may as well be defined as default, and
in that case the alphabet attribute to be optional). If there
are other values currently allowed, what are they? How are
they defined?
2.1.5 'alphabet' may not be the best name. Alphabets are sets of
[43] characters, usually with an ordering. The same set of characters
could be used in totally different notations.
2.1.5 What are the interactions of <phoneme> for foreign language
[44] segments? Do processors have to handle all of IPA, or only the
phonemes that are used in a particular language? Please clarify.
2.1.5, 1st example: Please try to avoid character entities, as it
[45] suggests strongly that this is the normal way to input this stuff.
(see also issue about utf-8 vs. iso-8859-1)
2.1.5 and 2.1.6: The 'alias' and 'ph' attributes in some
[46] cases will need additional markup (e.g. for fine-grained
prosody, but also for additional emphasis, bidirectionality).
This would also help tools for translation,...
But markup is not possible for attributes. These attributes
should be changed to subelements, e.g. similar to the <desc>
element inside <audio>.
2.1.5 and 2.1.6: Can you specify a null string for the ph and alias
[47] attributes? This may be useful in mixed formats where the
pronunciation is given by another means, e.g. with ruby annotation.
2.1.6 The <sub> element may easily clash or be confused with <sub>
[48] in HTML (in particular because the specification seems to be
designed to allow combinations with other markup vocabularies
without using different namespaces). <sub> should be renamed,
e.g. to <subst>.
2.1.6 For abbreviations,... there are various cases. Please check
[49] that all the cases in
http://lists.w3.org/Archives/Member/w3c-i18n-ig/2002Mar/0064.html
are covered, and that the users of the spec know how to handle
them.
2.1.6, 1st para: "the specified text" ->
[50] "text in the alias attribute value".
2.2.1, between the tables: "If there is no voice available for the
[51] requested language ... select a voice ... same language but different
region..." I'm not sure this makes sense. I could understand that
if there is no en-UK voice you'd maybe go for an en-US voice - this
is a different DIALECT of English. If there are no Japanese voices
available for Japanese text, I'm not sure it makes sense to use an
English voice. What happens in this situation?
2.2.1 It should be mentioned that in some cases, it may make sense to have
[52] a short piece of e.g. 'fr' text in an 'en' text been spoken by
an 'en' text-to-speech converter (the way it's often done by
human readers) rather than to throw an error. This is quite
different for longer texts, where it's useless to bother an
user.
2.2.1: We wonder if there's a need for multiple voices (eg. A group of kids)
[53]
2.2.1, 2nd example: You should include some text here.
[54]
2.2.1 The 'age' attribute should explicitly state that the integer
[55] is years, not something else.
2.2.1 The variant attribute should say what it's index origin is
[56] (e.g. either starting at 0 or at 1)
2.2.1 attribute name: (in the long term,) it may be desirable to use
[57] an URI for voices, and to have some well-defined format(s)
for the necessary data.
2.2.1, first example (and many other places): The line break between
[58] the <voice> start tag and the text "It's fleece was white as snow."
will have negative effects on visual rendering.
(also, "It's" -> "Its")
2.2.1, description of priorities of xml:lang, name, variant,...:
[59] It would be better to describe this clearly as priorities,
i.e. to say that for voice selection, xml:lang has highest
priority,...
2.2.3 What about <break> inside a word (e.g. for long words such as
[60] German)? What about <break> in cases where words cannot
clearly be identified (no spaces, such as in Chinese, Japanese,
Thai). <break> should be allowed in these cases.
2.2.3 and 2.2.4: "x-high" and "x-low": the 'x-' prefix is part of
[61] colloquial English in many parts of the world, but may be
difficult to understand for non-native English speakers.
Please add an explanation.
2.2.4: Please add a note that customary pitch levels and
[62] pitch ranges may differ quite a bit with natural language, and that
"high",... may refer to different absolute pitch levels for different
languages. Example: Japanese has general much lower pitch range than
Chinese.
2.2.4, 'baseline pitch', 'pitch range': Please provide definition/
[63] short explanation.
2.2.4 'as a percent' -> 'as a percentage'
[64]
2.2.4 What is a 'semitone'? Please provide a short explanation.
[65]
2.2.4 In pitch contour, are white spaces allowed? At what places
[66] exactly? In "(0%,+20)(10%,+30%)(40%,+10)", I would propose
to allow whitespace between ')' and '(', but not elsewhere.
This has the benefit of minimizing syntactict differences
while allowing long contours to be formatted with line breaks.
2.2.4, bullets: Editorial nit. It may help the first time reader to
[67] mention that 'relative change' is defined a little further down.
2.2.4, 4th bullet: the speaking rate is set in words per minute.
[68] In many languages what constitutes a word is often difficult to
determine, and varies considerably in average length.
So there have to be more details to make this work interoperably
in different languages. Also, it seems that 'words per minute'
is a nominal rate, rather than exactly counting words, which
should be stated clearly. An much preferable alternative is to use
another metric, such as syllables per minute, which has less
unclarity (not
2.2.4, 5th bullet: If the default is 100.0, how do you make it
[69] louder given that the scale ranges from 0.0 to 100.0?
(or, in other words, is the default to always shout?)
2.2.4, Please state whether units such as 'Hz' are case-sensitive
[70] or case-insensitive. They should be case-sensitive, because
units in general are (e.g. mHz (milliHz) vs. MHz (MegaHz)).
2.3.3 Please provide some example of <desc>
[71]
3.1 Requiring an XML declaration for SSML when XML itself
[72] doesn't require an XML declaration leads to unnecessary
discrepancies. It may be very difficult to check this
with an off-the-shelf XML parser, and it is not reasonable
to require SSML implementations to write their own XML
parsers or modify an XML parser. So this requirement
should be removed (e.g. by saying that SSML requires an XML
declaration when XML requires it).
3.3, last paragraph before 'The lexicon element' subtitle:
[73] Please also say that the determination of
what is a word may be language-specific.
3.3 'type' attribute on lexicon element: What's this attribute used
[74] for? The media type will be determined from the document that
is found at the 'uri' URI, or not?
4.1 'synthesis document fragment' -> 'speech synthesis document fragment'
[75]
4.1 Conversion to stand-alone document: xml:lang should not
[76] be removed. It should also be clear whether content of
non-synthesis elements should be removed, or only the
markup.
4.4 'requirement for handling of languages': Maybe better to
[77] say 'natural languages', to avoid confusion with markup
languages. Clarification is also needed in the following
bullet points.
4.5 This should say that a user agent has to support at least
[78] one natural language.
App A: 'http://www.w3c.org/music.wav': W3C's Web site is www.w3.org.
[79] But this example should use www.example.org or www.example.com.
App B: 'synthesis DTD' -> 'speech synthesis DTD'
[80]
App D: Why does this mentions 'recording'? Please remove or explain.
[81]
App E: Please give a reference for the application to the IETF/IESG/IANA
[82] for the content type 'application/ssml+xml'.
App F: 'Support for other phoneme alphabets.': What's a 'phoneme alphabet'?
[83]
App F, last paragraph: 'Unfortunately, ... no standard for designating
[84] regions...': This should be worded differently. RFC 3066 provides
for the registration of arbitrary extensions, so that e.g.
en-gb-accent-scottish and en-gb-accent-welsh could be registered.
App F, bullet 3: I guess you already know that intonation
[85] requirements can vary considerably across languages, so you'll
need to cast your net fairly wide here.
App G: What is meant by 'input' and 'output' languages? This is the
[86] first time this terminology is used. Please remove or clarify.
App G: 'overriding the SSML Processor default language': There should
[87] be no such default language. An SSML Processor may only
support a single language, but that's different from
assuming a default language.
Regards, Martin.
Received on Friday, 31 January 2003 22:50:00 UTC