Wyniki wyszukiwania - BazTech

1

Analysis and categorization of the Rusyn language using the whisper model : demographic influences on linguistic convergence

Małecki Paweł

International Journal of Electronics and Telecommunications

|

2026

|

Vol. 72, No. 1

art. no. 5, s. 1--7

EN

The article presents a detailed linguistic analysis of the Rusyn language, focusing on its complex and evolving features, such as pronunciation, as well as individual, regional, and historical variabilities. The investigation employed an artificial neural network based on the OpenAI Whisper model to perform analysis and categorization. Although the Whisper model was trained on data from the majority of state official languages, it was not specifically trained with samples of the Rusyn language due to its niche and minority/ethnic status. Consequently, speech samples in Rusyn were classified according to the most closely related available labels, allowing for the assessment of linguistic similarity between Rusyn and other (mostly) Slavic languages. The study incorporated a diverse user base segmented by gender, age, and geographic location (Poland, Ukraine, Slovakia, Serbia), revealing significant resemblances to the dominant languages within these countries and demonstrating correlations between the computed linguistic similarity and the speakers’ age.

2

Improving the vowel classification accuracy using varying signal frame length

Gmyrek Stanislaw, Hossa Robert

Vibrations in Physical Systems

|

2025

|

Vol. 36, nr 1

art. no. 2025114

EN

The parts of speech influenced by glottal pulse excitation, the vocal tract, and the speaker's lips shape the voiced components of the speech signal. On the other hand, semantic information in speech is primarily shaped by the vocal tract. However, the irregularity of the glottal excitation's periodicity contributes to a significant dispersion of the parameterization coefficients, introducing fluctuations into the amplitude spectrum. This study proposes a technique to mitigate the impact of this irregularity on the feature vector. It involves using a variable signal frame length synchronized with the fundamental period 𝑇𝑇0 and averaging amplitude spectra over a single period to minimize noise effects, smooth out the characteristics, and reduce the estimator variance. By utilizing the derived HFCC parameters, statistical models representing individual Polish vowels were created using mixtures of Gaussian distributions. Additionally, the impact of these correction concepts on the classification accuracy of speech frames containing Polish vowels was examined.

3

The Influence of the Amplitude Spectrum Correction in the HFCC Parametrization on the Quality of Speech Signal Frame Classification

Gmyrek Stanislaw, Hossa Robert, Makowski Ryszard

Archives of Acoustics

|

2025

|

Vol. 50, nr 1

59--67

EN

The voiced parts of the speech signal are shaped by glottal pulse excitation, the vocal tract, and the speaker’s lips. Semantic information contained in speech is shaped mainly by the vocal tract. Unfortunately, the quasiperiodicity of the glottal excitation, in the case of HFCC parameterization, is one of the factors affecting the significant scatter of the feature vector values by introducing ripples into the amplitude spectrum. This paper proposes a method to reduce the effect of quasiperiodicity of the excitation on the feature vector. For this purpose, blind deconvolution was used to determine the vocal tract transfer function estimator and the corrective function of the amplitude spectrum. Then, on the basis of the obtained HFCC parameters, statistical models of individual Polish speech phonemes were developed in the form of mixtures of Gaussian distributions, and the influence of the correction on the quality of classification of speech frames containing Polish vowels was investigated. The aim of the correction was to narrow the GMM distributions, which, according to detection theory, reduces the classification errors. The results obtained confirm the effectiveness of the proposed method.

4

The Impact of Training Strategies on Overfitting in Vowel Classification Using PS-HFCC Parametrization for Automatic Speech Recognition

Gmyrek Stanislaw, Libal Urszula, Hossa Robert

Archives of Acoustics

|

2025

|

Vol. 50, nr 3

371--382

EN

This paper investigates the overfitting problem in vowel classification task for automatic speech recognition (ASR). It utilizes a pitch synchronized human factor cepstral coefficients (PS-HFCC) as the parametrization method, which outperforms traditional methods like HFCC and mel-frequency cepstral coefficients (MFCC) in frame-level classification accuracy. While deep learning models are prevalent in contemporary ASR systems, they often lack explainability, a characteristic of classical classifiers. Therefore, this study examines overfitting phenomenon using a range of classifiers with well-understood properties. Specifically, it analyzes the impact of different training strategies on classifier performance, comparing the susceptibility to overfitting of several widely used classifiers, including the Gaussian mixture model (GMM), a standard approach in speech recognition. The analysis of training strategies considers various data splitting methods: random, speaker-based, and cluster-based. Our analysis of training strategies highlights the crucial role of data splitting methods: while random splitting is commonly used, it can lead to inflated accuracy due to overfitting. We demonstrate that speaker-independent splitting, where the classifier is trained on one set of speakers and tested on a separate, unseen set, is essential for robust evaluation and for accurately assessing generalization to new speakers. Potentially, the resulting insights may inform the future development and training of more reliable ASR systems.

5

Acoustic model for the classification of Polish vowels

Pondel-Sycz Karolina

Vibrations in Physical Systems

|

2024

|

Vol. 35, nr 1

art. no. 2024101

EN

The study explored the performance of vowel recognition using an acoustic model built on Audio Fingerprint techniques [1]. The research compares the performance of Support Vector Machines (SVMs), Hidden Markov Models (HMMs), Artificial Neural Networks (ANNs) and k-Nearest Neighbours (k-NN) classifiers in the recognition of isolated and within-word vowels and investigates the importance of different types of acoustic speech features in this process. Temporal, spectral, cepstral, formant, LPC and perceptual features of speech were examined. Importance of features was tested using a random forest classifier. Vowel classification was tested at three confidence levels for feature importance: 90%, 95% and 99%. Two author databases consisting of a total of 1,200 samples from 20 speakers, recorded under household conditions, were used. The classifiers were evaluated by confusion matrix, accuracy, precision, sensitivity and F1 score. A segmentation of words into speech sounds was carried out using a tool based on BiLSTM recurrent neural networks and the BIC criterion. Three most important features were determined: power spectral density, spectral cut-off, and Power-Normalised Cepstral Coefficients. In the isolated vowel recognition task, the SVM classifier was the most effective with a feature significance confidence level of 95% obtaining accuracy = 81%, precision = 81%, sensitivity = 81%, F1 score = 80%. In the task of recognising a vowel within a word, it was verified if the algorithm detected the presence of vowels in the correct segment and if it recognised the correct vowel within it. The best results were obtained by the k-NN classifier (statistical confidence level of feature importance of 99.9%). However, these results were low, correct recognition of the vowel in the word: A, E, U: 20%, I, O: 7%, Y: 23%. This indicates strong influence of the neighbourhood of other speech sounds in speech on the acoustic model of vowels and their recognition.

6

End-To-End deep neural models for Automatic Speech Recognition for Polish Language

Pondel-Sycz Karolina, Pietrzak Agnieszka Paula, Szymla Julia

International Journal of Electronics and Telecommunications

|

2024

|

Vol. 70, No. 2

315--321

EN

This article concerns research on deep learning models (DNN) used for automatic speech recognition (ASR). In such systems, recognition is based on Mel Frequency Cepstral Coefficients (MFCC) acoustic features and spectrograms. The latest ASR technologies are based on convolutional neural networks (CNNs), recurrent neural networks (RNNs) and Transformers. The article presents an analysis of modern artificial intelligence algorithms adapted for automatic recognition of the Polish language. The differences between conventional architectures and ASR DNN End-To-End (E2E) models are discussed. Preliminary tests of five selected models (QuartzNet, FastConformer, Wav2Vec 2.0 XLSR, Whisper and ESPnet Model Zoo) on Mozilla Common Voice, Multilingual LibriSpeech and VoxPopuli databases are demonstrated. Tests were conducted for clean audio signal, signal with bandwidth limitation and degraded. The tested models were evaluated on the basis of Word Error Rate (WER).

7

Amplitude spectrum correction to improve speech signal classification quality

Gmyrek Stanislaw, Hossa Robert, Makowski Ryszard

International Journal of Electronics and Telecommunications

|

2024

|

Vol. 70, No. 3

569-574

EN

The speech signal can be described by three key elements: the excitation signal, the impulse response of the vocal tract, and a system that represents the impact of speech production through human lips. The primary carrier of semantic content in speech is primarily influenced by the characteristics of the vocal tract. Nonetheless, when it comes to parameterization coefficients, the irregular periodicity of the glottal excitation is a significant factor that leads to notable variations in the values of the feature vectors, resulting in disruptions in the amplitude spectrum with the appearance of ripples. In this study, a method is suggested to mitigate this phenomenon. To achieve this goal, inverse filtering was used to estimate the excitation and transfer functions of the vocal tract. Subsequently, using the derived parameterisation coefficients, statistical models for individual Polish phonemes were established as mixtures of Gaussian distributions. The impact of these corrections on the classification accuracy of Polish vowels was then investigated. The proposed modification of the parameterisation method fulfils the expectations, the scatter of feature vector values was reduced.

8

Zastosowanie automatycznego rozpoznawania mowy do wspomagania pracy sztygara prowadzącego zmianę w komorze maszyn ciężkich

Stefaniak Paweł, Koperska Wioletta, Stachowiak Maria, Śliwiński Paweł

Cuprum : czasopismo naukowo-techniczne górnictwa rud

|

2024

|

nr 1-2

17--30

PL

W artykule przedstawiono prototyp systemu rozpoznawania mowy zintegrowany z bazami systemów MES/ERP i technikami text mining, przystosowany do warunków i potrzeb ewidencyjnych sztygara prowadzącego zmianę w komorze maszyn ciężkich (KMC). Głównym celem rozwiązania jest odciążenie służb dozoru zmianowego w Oddziałach KGHM PM S.A. w zakresie rejestracji przebiegu realizacji prac w KMC, prowadzonej obecnie w elektronicznych systemach nadzoru operacyjnego i raportowania produkcji. Aktualnie w strukturze organizacyjnej KGHM funkcjonuje kilka specjalnych systemów, takich jak moduły SAP, obejmujące obszary HR, MM, PM/CMMS, oraz platforma eRaport. W ramach przeprowadzonych badań przeprowadzono inwentaryzację tych systemów, identyfikując wymagania funkcjonalne interfejsu wykorzystującego automatyczne rozpoznawanie mowy (ASR) do celów rejestracji wpisów, sporządzania notatek, przygotowywania korespondencji e-mail oraz pozyskiwania informacji niezbędnych do podejmowania bieżących decyzji. Określono również, które z procesów mogą być w pełni zautomatyzowane, a które jedynie w części. Opracowane wytyczne posłużyły do stworzenia oprogramowania, pełniącego funkcję asystenta głosowego, reagującego na precyzyjnie określony zakres komend, a także obsługującego specjalistyczny słownik i żargon kopalniany oraz wewnętrzną kodyfikację baz systemowych KGHM. Całość oprogramowania obejmuje aplikację mobilną dostępną dla urządzeń z systemem Android, a także skrypty w języku Python integrujące narzędzia uczenia maszynowego, analizy tekstu i rozpoznawania mowy. W celu oceny prototypowych rozwiązań, przeprowadzono szereg testów w kopalni i na powierzchni według opracowanych scenariuszy eksperymentalnych. Stworzono również mechanizm umożliwiający określenie tempa zdobywania doświadczenia przez użytkownika. Dodatkowo, przeprowadzona została ankieta wśród uczestników testów, która pozwoliła zidentyfikować obszary, możliwe do usprawnienia w ramach potencjalnego wdrożenia.

EN

A prototype speech recognition system integrated with MES/ERP databases and text mining techniques was developed in this study, tailored to the recording conditions and needs of the mine supervisor overseeing operations in heavy machinery chambers. The primary objective of the solution is to relieve the workload of the heavy equipment foreman by improving the recording of work progress in heavy machinery chambers, currently managed in electronic operational supervision and production reporting systems. Presently, several analogous systems operate within the organizational structure of KGHM, such as SAP modules covering HR, MM, PM/CMMS areas, and the eReport platform. Through a comprehensive examination of these systems, functional requirements for an interface utilizing automatic speech recognition (ASR) for recording entries, note-taking, email correspondence preparation, and retrieval of information necessary for ongoing decision-making were identified. The study also determined processes that can be fully automated and those only partially. The formulated guidelines played a pivotal role in shaping the development of software designed as a voice assistant. This software is highly responsive to strict defined set of commands, adept at handling mining industry terminology, and proficient in interfacing with the internal coding of KGHM's system databases. The software comprises a mobile application for Android devices and Python scripts integrating machine learning, text mining, and ASR tools. To evaluate the prototype solutions, a series of tests were conducted both underground and, on the surface, following defined experimental scenarios. A mechanism was established to determine the user's learning progress. Additionally, a survey among test participants was administered to identify areas for potential improvement in the context of prospective implementation.

9

A system dedicated to Polish automatic speech recognition – overview of solutions

Pondel-Sycz Karolina, Bilski Piotr

Bulletin of the Polish Academy of Sciences. Technical Sciences

|

2024

|

Vol. 72, nr 4

art. no. e149818

EN

The paper presents the analysis of modern Artificial Intelligence algorithms for the automated system supporting human beings during their conversation in Polish language. Their task is to perform Automatic Speech Recognition (ASR) and process it further, for instance fill the computer-based form or perform the Natural Language Processing (NLP) to assign the conversation to one of predefined categories. The State-of-the-Art review is required to select the optimal set of tools to process speech in the difficult conditions, which degrade accuracy of ASR. The paper presents the top-level architecture of the system applicable for the task. Characteristics of Polish language are discussed. Next, existing ASR solutions and architectures with the End-To-End (E2E) deep neural network (DNN) based ASR models are presented in detail. Differences between Recurrent Neural Networks (RNN), Convolutional Neural Networks (CNN) and Transformers in the context of ASR technology are also discussed.

10

Rozpoznawanie głosu i identyfikacja mówcy: przegląd wybranych metod rozpoznawania cech biometrycznych mowy

Śliwak-Orlicki Tomasz, Górski Krzysztof

Przegląd Elektrotechniczny

|

2023

|

R. 99, nr 9

225--229

PL

W niniejszej pracy przedstawiono ogólnie rozwój technologii rozpoznawania mowy, począwszy od pierwszych eksperymentów XIX wieku, aż po współczesne osiągnięcia w tej dziedzinie. Przeanalizowano przekształcenia technologiczne na przestrzeni ostatnich lat, omówiono kluczowe odkrycia oraz najważniejsze wydarzenia, które odegrały istotną rolę w rozwoju tej dziedziny, wskazując jednocześnie wybrane procesy wspomagające skuteczność rozpoznawania mowy pod kątem identyfikacji biometrycznej. Przedstawiono w zarysie charakterystyczne cechy wymowy dla języka polskiego.

EN

This paper presents a general overview of the development of speech recognition technology, from the first experiments of the 19th century to modern developments in this field. It analyses technological transformations over the past years, discusses key discoveries and key events that have played a significant role in the development of this field, while highlighting selected processes that support the effectiveness of speech recognition in terms of biometric identification. The characteristic features of pronunciation for the Polish language are outlined.

11

Continuous speech commands recognition with thai language used support vector machine technique: A case study of speech commands control for mobile robots

Muangjaroen Supavit, Udomsiri Sakol

Przegląd Elektrotechniczny

|

2023

|

R. 99, nr 6

181--187

EN

This article presents the application of an automatic speech recognition by continuous speech commands recognition with Thai language as a speaker verification model, this is a case study of speech commands control of mobile robots. The design of the automatic speech recognition system consisted of 3 steps: The first we analyzed the signal processing of the continuous speech commands and compared the accuracy of the speech recognition with a time frame adjustment and the overlapped period of signal filtered with the window function, The second we proceed to find the feature extraction of speech commands using format frequency techniques and configured the feature extraction with format frequencies of F1, F2, and F3,The last step was to design the recognition using Support Vector Machine technique to check the accuracy of an automatic speech recognition. These is support vector machine classification algorithm provides a comparison of the filtered function window and compares the accuracy of the time frame scaled and the overlapped time of the filtered, which gives different values of precision. From the experiment, the researcher found that are applied a Hanging function the test results of the test result of the "forward" speech commands has an accuracy of 81.92% but kind of Gaussian function the test results of the "backward" speech commands has an accuracy of 83.69%, the "turn left" speech commands had an accuracy of 82.81%, the "turn right" speech commands had an accuracy of 85.56% and the "Stop first" speech commands has an accuracy of 86.78% and speech recognition by continuous speech commands recognition with Thai language was applied an every function the test results of the overall performance of the speech commands has an accuracy of 83.88%.

PL

Artykuł przedstawia zastosowanie automatycznego rozpoznawania mowy poprzez ciągłe rozpoznawanie poleceń głosowych z językiem tajskim jako modelem weryfikacji mówiącego, jest to studium przypadku sterowania poleceniami głosowymi robotów mobilnych. Projekt systemu automatycznego rozpoznawania mowy składał się z 3 etapów: W pierwszym przeanalizowano przetwarzanie sygnału ciągłych poleceń głosowych i porównano dokładność rozpoznawania mowy z dopasowaniem przedziału czasowego i nakładającym się okresem sygnału filtrowanego funkcją okna, Następnie przystępujemy do znalezienia ekstrakcji funkcji poleceń głosowych przy użyciu technik formatowania częstotliwości i skonfigurowania ekstrakcji cech z częstotliwościami formatu F1, F2 i F3. Ostatnim krokiem było zaprojektowanie rozpoznawania przy użyciu techniki maszyny wektorów nośnych w celu sprawdzenia dokładności automatyczne rozpoznawanie mowy. Jest to algorytm klasyfikacji maszyny wektorów nośnych, który zapewnia porównanie przefiltrowanego okna funkcji i porównuje dokładność skalowanych ram czasowych oraz nakładających się czasów filtrowanych, co daje różne wartości precyzji. Na podstawie eksperymentu badacz odkrył, że po zastosowaniu funkcji wiszącej wyniki testu wyników poleceń głosowych „do przodu” mają dokładność 81,92%, ale rodzaj funkcji Gaussa wyniki testu poleceń głosowych „wstecz” mają dokładność 81,92% dokładność 83,69%, polecenia głosowe „skręć w lewo” miały dokładność 82,81%, polecenia głosowe „skręć w prawo” miały dokładność 85,56%, a polecenia głosowe „Najpierw zatrzymaj” mają dokładność 86,78%, a rozpoznawanie mowy przez zastosowano ciągłe rozpoznawanie poleceń głosowych w języku tajskim, a wyniki testu ogólnej wydajności poleceń głosowych mają dokładność 83,88%.

12

Boosting conversational AI correctness by accounting for ASR errors using a sequence to sequence model

Jadczak Szymon, Jaworski Rafał

Annals of Computer Science and Information Systems

|

2023

|

Vol. 35

1325--1328

EN

This paper describes the winning submission to the challenge CAICCAIC: Center for Artificial Intelligence Challenge on Conversational AI Correctness. The aim of the challenge was to design a mechanism of natural language understanding capable of interpreting user prompts. The prompts were the output of an automatic speech recognition system and therefore contained errors. In this scenario, it was necessary to apply techniques of accounting for these errors. As per the results of the challenge, the most effective technique proved to be an original use of a sequence to sequence model.

13

BIGOS - Benchmark Intended Grouping of Open Speech Corpora for Polish Automatic Speech Recognition

Junczyk Michał

Annals of Computer Science and Information Systems

|

2023

|

Vol. 35

585--590

EN

This paper presents a Benchmark Intended Grouping of Open Speech (BIGOS), a new corpus designed for Polish Automatic Speech Recognition (ASR) systems. This initial version of the benchmark leverages 1,900 audio recordings from 71 distinct speakers, sourced from 10 publicly available speech corpora. Three proprietary ASR systems and five open-source ASR systems were evaluated on a diverse set of recordings and the corresponding original transcriptions. Interestingly, it was found that the performance of the latest open-source models is on par with that of more established commercial services. Furthermore, a significant influence of the model size on system accuracy was observed, as well as a decrease in scenarios involving highly specialized or spontaneous speech. The challenges of using public datasets for ASR evaluation purposes and the limitations based on this inaugural benchmark are critically discussed, along with recommendations for future research. BIGOS corpus and associated tools that facilitate replication and customization of the benchmark are made publicly available.

14

Preliminary Evaluation of Convolutional Neural Network Acoustic Model for Iban Language Using NVIDIA NeMo

Michael Steve Olsen, Juan Sarah Samson, Mit Edwin

Journal of Telecommunications and Information Technology

|

2022

|

nr 1

43--53

EN

For the past few years, artificial neural networks (ANNs) have been one of the most common solutions relied upon while developing automated speech recognition (ASR) acoustic models. There are several variants of ANNs, such as deep neural networks (DNNs), recurrent neural networks (RNNs), and convolutional neural networks (CNNs). A CNN model is widely used as a method for improving image processing performance. In recent years, CNNs have also been utilized in ASR techniques, and this paper investigates the preliminary result of an end-to-end CNN-based ASR using NVIDIA NeMo on the Iban corpus, an under-resourced language. Studies have shown that CNNs have also managed to produce excellent word error (WER) rates for the acoustic model on ASR for speech data. Conversely, results and studies concerned with under-resourced languages remain unsatisfactory. Hence, by using NVIDIA NeMo, a new ASR engine developed by NVIDIA, the viability and the potential of this alternative approach are evaluated in this paper. Two experiments were conducted: the number of resources used in the works of our ASR’s training was manipulated, as was the internal parameter of the engine used, namely the epochs. The results of those experiments are then analyzed and compared with the results shown in existing papers.

15

Automatic Speech Recognition and its Application to Media Monitoring

Cecko R., Jamroży J., Jęśko W., Kuśmierek E., Lange M., Owsianny M.

Computational Methods in Science and Technology

|

2021

|

Vol. 27, No. 2

41--55

EN

In this paper we present application of the automatic speech recognition technology in the area of media monitoring. We describe the use of computational models and methods by two ASR technologies, namely a Hidden Markov Model with a Gaussian Mixture Model and Deep Neural Networks, that were crucial in the ASR development. Both approaches were implemented in our speech recognition ARM-1 engine developed for the Polish language. We provide details on the implementation choices, specifically adjustments made for media monitoring application guided by the characteristics of media content. Performance of both versions of our engine is evaluated and compared.

16

Hybrid CNN-Ligru acoustic modeling using sincnet raw waveform for hindi ASR

Kumar Ankit, Aggarwal Rajesh Kumar

Computer Science

|

2020

|

T. 21 (4)

397-417

EN

Deep neural networks (DNN) currently play a most vital role in automatic speech recognition (ASR). The convolution neural network (CNN) and recurrent neural network (RNN) are advanced versions of DNN. They are right to deal with the spatial and temporal properties of a speech signal, and both properties have a higher impact on accuracy. With its raw speech signal, CNN shows its superiority over precomputed acoustic features. Recently, a novel first convolution layer named SincNet was proposed to increase interpretability and system performance. In this work, we propose to combine SincNet-CNN with a light-gated recurrent unit (LiGRU) to help reduce the computational load and increase interpretability with a high accuracy. Different configurations of the hybrid model are extensively examined to achieve this goal. All of the experiments were conducted using the Kaldi and Pytorch-Kaldi toolkit with the Hindi speech dataset. The proposed model reports an 8.0% word error rate (WER).

17

Allophones in automatic speech recognition

Kozierski P., Sadalla T., Dąbrowski A., Giernacki W.

Studia z Automatyki i Informatyki

|

2016

|

T. 41

47--53

EN

The common approach to the speech recognition problem is the use of phonemes as basic parts of speech. The authors proposed allophones usage instead. For rarer allophones the conversion into other allophones (4 selection methods) has been proposed. Based on the obtained results one can say that effective use of the additional information from allophonic notation will not be possible without modification of currently used algorithms.

PL

Typowym podejściem do zagadnienia rozpoznawania mowy jest branie pod uwagę fonemów, jako podstawowych części mowy. Zamiast tego autorzy zaproponowali wykorzystanie alofonów. Dla najrzadziej występujących alofonów zaproponowano ich zamianę na inne alofony – zaproponowano 4 metody wyboru głosek do zamiany. Na podstawie uzyskanych wyników stwierdzono, że efektywne wykorzystanie dodatkowych informacji, jakie niosą alofony, nie będzie możliwe bez modyfikacji obecnie dostępnych algorytmów.

18

System rozpoznawania mowy polskiej dla robota społecznego

Zygadło A., Janicki A., Dąbek P.

Pomiary Automatyka Robotyka

|

2016

|

R. 20, nr 4

27--36

PL

W artykule przedstawiono system automatycznego rozpoznawania mowy polskiej dedykowany dla robota społecznego. System oparty jest na bezpłatnej i otwartej bibliotece oprogramowania pocketsphinx (CMU Sphinx). Przygotowano zbiory nagrań: treningowy i testowy wraz z transkrypcjami. Zbiór treningowy obejmował głosy 10 kobiet i 10 mężczyzn i został przygotowany na podstawie audiobooków, natomiast zbiór testowy – głosy 3 kobiet i 3 mężczyzn nagrane w warunkach laboratoryjnych specjalnie na potrzeby pracy. Przygotowany zbiór fonemów dla języka polskiego, składający się z 39 fonemów, opracowany został na podstawie dwóch popularnych zbiorów dostępnych danych. Słownik fonetyczny opracowano za pomocą funkcjonalności konwersji grapheme-to-phoneme z biblioteki eSpeak. Model statystyczny języka dla tekstu referencyjnego składającego się z 76 komend wygenerowano za pomocą programu cmuclmtk (CMU Sphinx). Uczenie modelu akustycznego oraz test jakości rozpoznawania mowy przeprowadzono za pomocą programu sphinxtrain (CMU Sphinx). W warunkach laboratoryjnych uzyskano wskaźnik błędu rozpoznawania słów (WER) na poziomie 4% i błędu rozpoznawania zdań (SER) na poziomie 9%. Przeprowadzono też badania systemu w warunkach rzeczywistych na grupie testowej złożonej z 2 kobiet i 3 mężczyzn, uzyskując wstępne wyniki rozpoznawania na poziomie 10% (SER) z bliskiej odległości oraz 60% (SER) z odległości 3 m. Określono kierunki dalszych prac.

EN

Automatic Speech Recognition system for Polish and dedicated for social robotics applications is presented. The system is based on free and open software library pocketsphinx (CMU Sphinx). Training and test databases were prepared with transcriptions; the training database comprised voices of 10 women and 10 men, and it was prepared based on audiobooks, whereas the test database comprised voices of 3 women and 3 men recorded in laboratory conditions as a part of the present work. A phoneme set for Polish consisting of 39 phonemes based on two popular sets from other researchers was prepared. The phonetic dictionary was obtained using graphemeto-phoneme conversion from the eSpeak tool for speech synthesis. The language statistic model for the reference text including 76 commands was generated using cmuclmtk tool (CMU Sphinx). Training of the acoustic model and test of quality of speech recognition was conducted using the sphinxtrain tool (CMU Sphinx). The following error rates were obtained for laboratory conditions: 4% (WER) and 9% (SER). Next, investigations of the system in relevant real environment were conducted. The initial, tentative results are about 10% (SER) for the close distance of a speaker to a microphone, and about 60% (SER) for 3 m speaker-microphone distance. Directions of future works are formulated.

19

Grupowanie mówców i jego skuteczność dla języka polskiego

Zambrzycka A., Makowski R., Hossa R.

Elektronika : konstrukcje, technologie, zastosowania

|

2016

|

Vol. 57, nr 7

45--50

PL

Grupowanie mówców w zbiory o podobnych cechach akustycznych ich mowy, obok normalizacji i adaptacji, jest skuteczną metodą poprawy jakości systemów automatycznego rozpoznawania mowy. W pracy przedstawiono metody grupowania, dla których punktem wyjścia jest model akustyczny wszystkich mówców oraz ich efektywność dla mowy polskiej w odniesieniu głównie do samogłosek. Rozwiązania te okazały się być skuteczne nawet przy wykorzystaniu superkrótkiej wypowiedzi. Uzyskana poprawa jakości rozpoznawania ramek mierzona za pomocą frame error rate wynosi około 4%.

EN

Clustering of speakers into groups of similar acoustic features is, besides for normalization and adaptation, an efficient method of improving the quality of systems of automatic speech recognition. New approaches of speaker clustering based on the acoustic model for all speakers and their efficiency for Polish speech, mostly regarding vowels, are presented and discussed in this paper. Results show the strong performance of the new solutions, even when super short speech segments were used. The obtained quality improvement of frame recognition measured by frame error rate was about 4%.

20

An Effective Speaker Clustering Method using UBM and Ultra-Short Training Utterances

Hossa R., Makowski R.

Archives of Acoustics

|

2016

|

Vol. 41, No. 1

107--118

EN

The same speech sounds (phones) produced by different speakers can sometimes exhibit significant differences. Therefore, it is essential to use algorithms compensating these differences in ASR systems. Speaker clustering is an attractive solution to the compensation problem, as it does not require long utterances or high computational effort at the recognition stage. The report proposes a clustering method based solely on adaptation of UBM model weights. This solution has turned out to be effective even when using a very short utterance. The obtained improvement of frame recognition quality measured by means of frame error rate is over 5%. It is noteworthy that this improvement concerns all vowels, even though the clustering discussed in this report was based only on the phoneme a. This indicates a strong correlation between the articulation of different vowels, which is probably related to the size of the vocal tract.