Pronunciation generation, non english

[Li3]

Hello,

With my little knowledge of Sphinx tool set, I understand that for english
language, Sphinx has a predefined 40 phone set and provides online tools to
generate pronunciation dictionaries (for given set of words) based on this
phone set and its proprietary english dictionary.

Wondering how one needs to go about deciding a phone set and generating a
pronunciation dictionary (in Sphinx format) for any other language so that
he/she could use Pocketsphinx for doing speech recognition in that language.
(FSG based. No language model).

thanks, Li

[Nickolay V. Shmyrev]

provides online tools to generate pronunciation dictionaries (for given set
of words) based

Offline tools are also provided

on this phone set and its proprietary english dictionary.

Dictionary is open

http://www.speech.cs.cmu.edu/cgi-bin/cmudict

Wondering how one needs to go about deciding a phone set and generating a
pronunciation dictionary (in Sphinx format) for any other language so that
he/she could use Pocketsphinx for doing speech recognition in that language.
(FSG based. No language model).

See

http://cmusphinx.sourceforge.net/wiki/tutorialdict

[Joseph S. Wisniewski]

Look for existing phone sets in multi-language language models for Sphinx on
this site

• CMU Chinese

• LIUM French

• Mexican Spanish, origin unknown

• VoxForge Dutch

• VoxForge Russian

• VoxForge German

• VoxForge Spanish

Also look at the VoxForfge site, where there's additional models and phone
sets for

• Italian

• Hebrew

• Greek

• Bulgarian

• Portuguese

• French, with a different phone set than LIUM, I believe

The Julius site for Japaneses
The iAtros site for Spanish that, I believe, is different than the VoxForge
set.

Then look at the speech synthesis sites like Festival and Flite, to see what
phone sets they use for different languages. That's largely what VoxForge did,
using phone sets and dictionaries from speech synthesis projects.

[Li3]

Thanks for pointers wiz, Li

[vera]

hello.. please help me..
I try to build speech to text Japanese version with Pocketsphinx but I can't
make a dump file, so I can't train it.
This is the sample of my corpus text.
あ あおい あかい あした あつい

あなた あの ありがとう い いいえ

あのひとはだれですか いきます いくら

いくらですか いただきます いち いちがつ いつ

いま う え お おいしい おおい おげん きですか

おなか おなかすいた おねがいします

おはよう おやすみ か かぎ かさ かぜ

When I compile the sentences at http://www.speech.cs.cmu.edu/tools/lmtool-
new.html

there are errors in .log_pronounce
WARN> found an empty line
WARN> found an empty line
WARN> found an empty line
WARN> found an empty line
WARN> found an empty line
WARN> found an empty line
WARN> found an empty line
WARN> found an empty line
WARN> found an empty line
WARN> found an empty line
WARN> found an empty line
WARN> found an empty line

I don't know if the sphinx knowledge base tool can't generate Japanese
characters. But Pocketsphinx has mandarin version so I think I can make
Japanese version.

What should I do?

[Nickolay V. Shmyrev]

Online lmtool works only for English. See the tutorial for other options

http://cmusphinx.sourceforge.net/wiki/tutoriallm

[vera]

nshmyrev, thanks for your answer..
Is it possible that julius (japanese open source) can be combined with
pocketsphinx? Do you know online lmtool for Japanese? Pocketsphinx can only
use Arpa system, is that right?

[Nickolay V. Shmyrev]

Is it possible that julius (japanese open source) can be combined with
pocketsphinx?

No

Do you know online lmtool for Japanese?

No

Pocketsphinx can only use Arpa system, is that right?

Pocketsphinx can use any trigram language model in ARPA format. Model, not
system.