Merge pull request #285 from Holzhaus/julius-stt-engine

Added basic Julius STT engine and vocabcompiler

2 files changed, 244 insertions, 1 deletions

diff --git a/client/stt.py b/client/stt.py
index 95ed487..c192e6a 100644
--- a/client/stt.py
+++ b/client/stt.py
@@ -7,6 +7,8 @@ import tempfile
 import logging
 import urllib
 import urlparse
+import re
+import subprocess
 from abc import ABCMeta, abstractmethod
 import requests
 import yaml
@@ -176,6 +178,97 @@ class PocketSphinxSTT(AbstractSTTEngine):
         return diagnose.check_python_import('pocketsphinx')
 
 
+class JuliusSTT(AbstractSTTEngine):
+    """
+    A very basic Speech-to-Text engine using Julius.
+    """
+
+    SLUG = 'julius'
+    VOCABULARY_TYPE = vocabcompiler.JuliusVocabulary
+
+    def __init__(self, vocabulary=None, hmmdefs="/usr/share/voxforge/julius/" +
+                 "acoustic_model_files/hmmdefs", tiedlist="/usr/share/" +
+                 "voxforge/julius/acoustic_model_files/tiedlist"):
+        self._logger = logging.getLogger(__name__)
+        self._vocabulary = vocabulary
+        self._hmmdefs = hmmdefs
+        self._tiedlist = tiedlist
+        self._pattern = re.compile(r'sentence(\d+): <s> (.+) </s>')
+
+        # Inital test run: we run this command once to log errors/warnings
+        cmd = ['julius',
+               '-input', 'stdin',
+               '-dfa', self._vocabulary.dfa_file,
+               '-v', self._vocabulary.dict_file,
+               '-h', self._hmmdefs,
+               '-hlist', self._tiedlist,
+               '-forcedict']
+        cmd = [str(x) for x in cmd]
+        self._logger.debug('Executing: %r', cmd)
+        with tempfile.SpooledTemporaryFile() as out_f:
+            with tempfile.SpooledTemporaryFile() as f:
+                with tempfile.SpooledTemporaryFile() as err_f:
+                    subprocess.call(cmd, stdin=f, stdout=out_f, stderr=err_f)
+            out_f.seek(0)
+            for line in out_f.read().splitlines():
+                line = line.strip()
+                if len(line) > 7 and line[:7].upper() == 'ERROR: ':
+                    if not line[7:].startswith('adin_'):
+                        self._logger.error(line[7:])
+                elif len(line) > 9 and line[:9].upper() == 'WARNING: ':
+                    self._logger.warning(line[9:])
+                elif len(line) > 6 and line[:6].upper() == 'STAT: ':
+                    self._logger.debug(line[6:])
+
+    @classmethod
+    def get_config(cls):
+        # FIXME: Replace this as soon as we have a config module
+        config = {}
+        # HMM dir
+        # Try to get hmm_dir from config
+        profile_path = jasperpath.config('profile.yml')
+
+        if os.path.exists(profile_path):
+            with open(profile_path, 'r') as f:
+                profile = yaml.safe_load(f)
+                if 'julius' in profile:
+                    if 'hmmdefs' in profile['julius']:
+                        config['hmmdefs'] = profile['julius']['hmmdefs']
+                    if 'tiedlist' in profile['julius']:
+                        config['tiedlist'] = profile['julius']['tiedlist']
+        return config
+
+    def transcribe(self, fp, mode=None):
+        cmd = ['julius',
+               '-quiet',
+               '-nolog',
+               '-input', 'stdin',
+               '-dfa', self._vocabulary.dfa_file,
+               '-v', self._vocabulary.dict_file,
+               '-h', self._hmmdefs,
+               '-hlist', self._tiedlist,
+               '-forcedict']
+        cmd = [str(x) for x in cmd]
+        self._logger.debug('Executing: %r', cmd)
+        with tempfile.SpooledTemporaryFile() as out_f:
+            with tempfile.SpooledTemporaryFile() as err_f:
+                subprocess.call(cmd, stdin=fp, stdout=out_f, stderr=err_f)
+            out_f.seek(0)
+            results = [(int(i), text) for i, text in
+                       self._pattern.findall(out_f.read())]
+        transcribed = [text for i, text in
+                       sorted(results, key=lambda x: x[0])
+                       if text]
+        if not transcribed:
+            transcribed.append('')
+        self._logger.info('Transcribed: %r', transcribed)
+        return transcribed
+
+    @classmethod
+    def is_available(cls):
+        return diagnose.check_executable('julius')
+
+
 class GoogleSTT(AbstractSTTEngine):
     """
     Speech-To-Text implementation which relies on the Google Speech API.
diff --git a/client/vocabcompiler.py b/client/vocabcompiler.py
index ed9f425..0f8a648 100644
--- a/client/vocabcompiler.py
+++ b/client/vocabcompiler.py
@@ -8,7 +8,13 @@ import os
 import tempfile
 import logging
 import hashlib
+import subprocess
+import tarfile
+import re
+import contextlib
+import shutil
 from abc import ABCMeta, abstractmethod, abstractproperty
+import yaml
 
 import brain
 import jasperpath
@@ -325,6 +331,151 @@ class PocketsphinxVocabulary(AbstractVocabulary):
                     f.write(line)
 
 
+class JuliusVocabulary(AbstractVocabulary):
+    class VoxForgeLexicon(object):
+        def __init__(self, fname, membername=None):
+            self._dict = {}
+            self.parse(fname, membername)
+
+        @contextlib.contextmanager
+        def open_dict(self, fname, membername=None):
+            if tarfile.is_tarfile(fname):
+                if not membername:
+                    raise ValueError('archive membername not set!')
+                tf = tarfile.open(fname)
+                f = tf.extractfile(membername)
+                yield f
+                f.close()
+                tf.close()
+            else:
+                with open(fname) as f:
+                    yield f
+
+        def parse(self, fname, membername=None):
+            pattern = re.compile(r'\[(.+)\]\W(.+)')
+            with self.open_dict(fname, membername=membername) as f:
+                for line in f:
+                    matchobj = pattern.search(line)
+                    if matchobj:
+                        word, phoneme = [x.strip() for x in matchobj.groups()]
+                        if word in self._dict:
+                            self._dict[word].append(phoneme)
+                        else:
+                            self._dict[word] = [phoneme]
+
+        def translate_word(self, word):
+            if word in self._dict:
+                return self._dict[word]
+            else:
+                return []
+
+    PATH_PREFIX = 'julius-vocabulary'
+
+    @property
+    def dfa_file(self):
+        """
+        Returns:
+            The path of the the julius dfa file as string
+        """
+        return os.path.join(self.path, 'dfa')
+
+    @property
+    def dict_file(self):
+        """
+        Returns:
+            The path of the the julius dict file as string
+        """
+        return os.path.join(self.path, 'dict')
+
+    @property
+    def is_compiled(self):
+        return (super(self.__class__, self).is_compiled and
+                os.access(self.dfa_file, os.R_OK) and
+                os.access(self.dict_file, os.R_OK))
+
+    def _get_grammar(self, phrases):
+        return {'S': [['NS_B', 'WORD_LOOP', 'NS_E']],
+                'WORD_LOOP': [['WORD_LOOP', 'WORD'], ['WORD']]}
+
+    def _get_word_defs(self, lexicon, phrases):
+        word_defs = {'NS_B': [('<s>', 'sil')],
+                     'NS_E': [('</s>', 'sil')],
+                     'WORD': []}
+
+        words = []
+        for phrase in phrases:
+            if ' ' in phrase:
+                for word in phrase.split(' '):
+                    words.append(word)
+            else:
+                words.append(phrase)
+
+        for word in words:
+            for phoneme in lexicon.translate_word(word):
+                word_defs['WORD'].append((word, phoneme))
+        return word_defs
+
+    def _compile_vocabulary(self, phrases):
+        prefix = 'jasper'
+        tmpdir = tempfile.mkdtemp()
+
+        lexicon_file = jasperpath.data('julius-stt', 'VoxForge.tgz')
+        lexicon_archive_member = 'VoxForge/VoxForgeDict'
+        profile_path = jasperpath.config('profile.yml')
+        if os.path.exists(profile_path):
+            with open(profile_path, 'r') as f:
+                profile = yaml.safe_load(f)
+                if 'julius' in profile:
+                    if 'lexicon' in profile['julius']:
+                        lexicon_file = profile['julius']['lexicon']
+                    if 'lexicon_archive_member' in profile['julius']:
+                        lexicon_archive_member = \
+                            profile['julius']['lexicon_archive_member']
+
+        lexicon = JuliusVocabulary.VoxForgeLexicon(lexicon_file,
+                                                   lexicon_archive_member)
+
+        # Create grammar file
+        tmp_grammar_file = os.path.join(tmpdir,
+                                        os.extsep.join([prefix, 'grammar']))
+        with open(tmp_grammar_file, 'w') as f:
+            grammar = self._get_grammar(phrases)
+            for definition in grammar.pop('S'):
+                f.write("%s: %s\n" % ('S', ' '.join(definition)))
+            for name, definitions in grammar.items():
+                for definition in definitions:
+                    f.write("%s: %s\n" % (name, ' '.join(definition)))
+
+        # Create voca file
+        tmp_voca_file = os.path.join(tmpdir, os.extsep.join([prefix, 'voca']))
+        with open(tmp_voca_file, 'w') as f:
+            for category, words in self._get_word_defs(lexicon,
+                                                       phrases).items():
+                f.write("%% %s\n" % category)
+                for word, phoneme in words:
+                    f.write("%s\t\t\t%s\n" % (word, phoneme))
+
+        # mkdfa.pl
+        olddir = os.getcwd()
+        os.chdir(tmpdir)
+        cmd = ['mkdfa.pl', str(prefix)]
+        with tempfile.SpooledTemporaryFile() as out_f:
+            subprocess.call(cmd, stdout=out_f, stderr=out_f)
+            out_f.seek(0)
+            for line in out_f.read().splitlines():
+                line = line.strip()
+                if line:
+                    self._logger.debug(line)
+        os.chdir(olddir)
+
+        tmp_dfa_file = os.path.join(tmpdir, os.extsep.join([prefix, 'dfa']))
+        tmp_dict_file = os.path.join(tmpdir, os.extsep.join([prefix, 'dict']))
+        shutil.move(tmp_dfa_file, self.dfa_file)
+        shutil.move(tmp_dict_file, self.dict_file)
+
+        shutil.rmtree(tmpdir)
+
+
 def get_phrases_from_module(module):
     """
     Gets phrases from a module.
@@ -373,7 +524,6 @@ def get_all_phrases():
     return sorted(list(set(phrases)))
 
 if __name__ == '__main__':
-    import shutil
     import argparse
 
     parser = argparse.ArgumentParser(description='Vocabcompiler Demo')