This article includes the results of testing an existing speech recognition systems, using the Russian language model and dictionary. Objects of the research are question-answering systems, call-center processes, methods and systems for language processing. Several problems with Russian language speech recognition were identified and studied.
Introduction
In the research, algorithms for speech recognition are considered. These algorithmsare used by the following systems
for speech recognition: Sphinx [
Currently, CMU Sphinx includes language models for the several languages: English, Russian, German, Chinese and others. It gives an opportunity to create acoustic models for other languages. The project uses BSD license, which allows distributing the product commercially, the system also includes tools for speech recognition (keyword definition, pronunciation evaluation). The development is still in progress.
Yandex.Speech Kit – is a complex of speech technologies, developed by Yandex Company, it includes speech
recognition and synthesis. Speech Kit is used as the cloud service Speech Kit Cloud and Speech Kit Mobile [
Interaction with Speech Kit Cloud is performs HTTP API and includes different functions: 1. Interactive Voice Response 2. Automatic calls to transmit information about new services, to confirm an order or delivery, to remind about the record, the collection of meter readings. 3. Inquiries by phone without operator's participation, recording for reception and maintenance. 4. The voice interface of "smart house" systems. 5. The voice interface of robots.
6. Site management, using voice.
Speech Kit Cloud is designed to recognize small speech fragments about 30 seconds long [
Speech Kit Mobile SDK [
Deploying process of the speech recognition systems for testing the methods and algorithms Deploying CMU Sphinx and Google Speech Recognition.
To deploy the Pocket Sphinx system and Google Speech Recognition, a solution from a third-party developer was used, it is represented as a Speech Recognition Python package. It can be installed using the Pip tool: pip install SpeechRecognition pip install pocketsphinx The installation requires Python 3.3 and later, Pip. The Python wrapper package for the Pocket Sphinx system must be installed as well: We used Python 3.6. Installation was performed in the Windows terminal from PyCharm (using virtual environment) (Figure 1). https://sourceforge.net/projects/cmusphinx/files/Acoustic%20and%20Language%20Models/Russian/ Files of one of the downloaded models should be placed in the pocket sphinx-data folder in the folder of the speech_recognition package (Figure 2). After that, the launch of the following Python script will display the result of speech recognition in both systems (the data from the pre-recorded audio file is recognized - Figure 3).
The GitHub repository of the SpeechRecognition package and installation instructions: https://github.com/Uberi/speech_recognition In order to work with the system, a key must be obtained on the site: https://developer.tech.yandex.ru/. The test key (Figure.4) only works for a month and has a limit on the number of requests.
Speech Kit Cloud API is a WebAPI, which means that in order to work with the system, a POST request
should be sent, and then an XML file with the recognition results will processed (Figures 5-6). Instructions for
interacting with Speech Kit Cloud API are presented in [
import urllib.request as req from xml.dom import minidom key = '1e692527-ad23-4fdb-b463-b34e545f9a13' uuid = 'aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaab' file = 'english.wav' defread_file(): with open(file, 'rb') as f: return f.read() defget_text_xml(data):
request = req.Request(url='https://asr.yandex.net/asr_xml?uuid=' + uuid + '&key=' + key + '&topic=queries', headers={'Content-Type': 'audio/x-wav', 'Content-Length': len(data)})
response = req.urlopen(url=request, data=data) return response.read() defparse_xml(xml_string): xmldoc = minidom.parseString(xml_string) if xmldoc.getElementsByTagName('recognitionResults')[0].attributes['success'].value == '1': return xmldoc.getElementsByTagName('variant')[0].childNodes[0].nodeValue binary = read_file() xml = get_text_xml(binary) best_variant = parse_xml(xml) print(best_variant) 2. Research and testing of the speech recognition methods and algorithms This section presents the results of testing speech recognition methods and algorithms using existing systems Sphinx Speech Recognition, Yandex Speech Kit and Google Speech Recognition.
Fragments of the testing results for the Sphinx Speech Recognition system are presented in Tables 1-4.It includes only cases with errors.
Пылесос Тостер
да вы отсос то вздыхать Весы напольные везде на вольная пылесос весы напольные 0 пылесос весы напольные 0 Based on the information presented above, we can conclude that the Yandex system is good at recognizing short expressive phrases, as well as numerals. On contrary, the Google API is good at recognizing long phrases and terms. The Sphinx system is struggling with recognition of Russian speech. The results of this research are being used in development of the automatic system that calls the customers of TWIN.
This work is supported by Act 211 Government of the Russian Federation, contract № 02.A03.21.0006.