Automatic Speech Recognition Performance Research Articles

Whisper-SV: Adapting Whisper for low-data-resource speaker verification

Trained on 680,000 h of massive speech data, Whisper is a multitasking, multilingual speech foundation model demonstrating superior performance in automatic speech recognition, translation, and language identification. However, its applicability in speaker verification (SV) tasks remains unexplored, particularly in low-data-resource scenarios where labeled speaker data in specific domains are limited. To fill this gap, we propose a lightweight adaptor framework to boost SV with Whisper, namely Whisper-SV. Given that Whisper is not specifically optimized for SV tasks, we introduce a representation selection module to quantify the speaker-specific characteristics contained in each layer of Whisper and select the top-k layers with prominent discriminative speaker features. To aggregate pivotal speaker-related features while diminishing non-speaker redundancies across the selected top-k distinct layers of Whisper, we design a multi-layer aggregation module in Whisper-SV to integrate multi-layer representations into a singular, compacted representation for SV. In the multi-layer aggregation module, we employ convolutional layers with shortcut connections among different layers to refine speaker characteristics derived from multi-layer representations from Whisper. In addition, an attention aggregation layer is used to reduce non-speaker interference and amplify speaker-specific cues for SV tasks. Finally, a simple classification module is used for speaker classification. Experiments on VoxCeleb1, FFSVC, and IMSV datasets demonstrate that Whisper-SV achieves EER/minDCF of 2.22%/0.307, 6.14%/0.488, and 7.50%/0.582, respectively, showing superior performance in low-data-resource SV scenarios.

Automatic speech recognition using advanced deep learning approaches: A survey

Recent advancements in deep learning (DL) have posed a significant challenge for automatic speech recognition (ASR). ASR relies on extensive training datasets, including confidential ones, and demands substantial computational and storage resources. Enabling adaptive systems improves ASR performance in dynamic environments. DL techniques assume training and testing data originate from the same domain, which is not always true. Advanced DL techniques like deep transfer learning (DTL), federated learning (FL), and deep reinforcement learning (DRL) address these issues. DTL allows high-performance models using small yet related datasets, FL enables training on confidential data without dataset possession, and DRL optimizes decision-making in dynamic environments, reducing computation costs.This survey offers a comprehensive review of DTL, FL, and DRL-based ASR frameworks, aiming to provide insights into the latest developments and aid researchers and professionals in understanding the current challenges. Additionally, Transformers, which are advanced DL techniques heavily used in proposed ASR frameworks, are considered in this survey for their ability to capture extensive dependencies in the input ASR sequence. The paper starts by presenting the background of DTL, FL, DRL, and Transformers and then adopts a well-designed taxonomy to outline the state-of-the-art (SOTA) approaches. Subsequently, a critical analysis is conducted to identify the strengths and weaknesses of each framework. Additionally, a comparative study is presented to highlight the existing challenges, paving the way for future research opportunities.

A closer look at reinforcement learning-based automatic speech recognition

Reinforcement learning (RL) has demonstrated effectiveness in improving model performance and robustness for automatic speech recognition (ASR) tasks. Researchers have employed RL-based training strategies to enhance performance beyond conventional supervised or semi-supervised learning. However, existing approaches treat RL as a supplementary tool, leaving the untapped potential of RL training largely unexplored. In this paper, we formulate a novel pure RL setting where an ASR model is trained exclusively through RL via human feedback metrics, e.g., word error rate (WER) or binary reward. This approach promises to significantly simplify the annotation process if we could replace the conventional onerous annotation with a single numeric value in a human–computer interaction (HCI) way. Our experiments demonstrate the feasibility of this new setting and also identify two main inherent issues in conventional RL-based ASR training that may lead to performance degradation: (1) the mismatch issue between the action and reward has been commonly overlooked in Connectionist Temporal Classification (CTC) based models, which is attributed to the inherent CTC alignment mapping issue; (2) the classic exploration–exploitation trade-off still exists in the sampling stage of RL-based ASR, and finding the balance between them becomes a challenge. To address these issues, we first propose a new RL-based approach named CTC-aligned Policy Gradient (CTC-PG), which provides a unified formulation for different sampling strategies and alleviates the mismatch issue of action and reward in CTC-based models. Moreover, we propose Focal Sampling to balance the trade-off between exploration and exploitation with a flexible temperature parameter. Experiment results on LibriSpeech dataset showcase the effectiveness and robustness of our methods by harnessing the full potential of RL in training ASR models.

Keep the noise down: On the performance of automatic speech recognition of voice-recordings in web surveys

Voice-recordings are increasingly implemented in web surveys, but the resulting audio data need to be transcribed before analysis. Since manual coding is too time- and work-intensive, researchers often rely on automatic speech recognition (ASR) systems for the transcription of the voice-recordings. However, ASR tools might create partly incorrect transcriptions and potentially change the content of responses. If the ASR performance (i.e., accuracy and validity) differs by subgroup and contextual factors, a bias is introduced in the analysis of open-ended questions. We assessed the impact of sociodemographic and contextual factors on the accuracy and validity of ASR transcriptions with data from the Longitudinal Internet Studies for the Social Sciences (LISS) panel collected in December 2020. We find that background noise reduces the accuracy and validity of ASR transcriptions. In addition, validity improved when the respondent was alone during the survey. Fortunately, we did not find any evidence of systematic differences across subgroups (age, sex, education), devices or respondent location.

Improving Dysarthric Speech Segmentation With Emulated and Synthetic Augmentation.

Acoustic features extracted from speech can help with the diagnosis of neurological diseases and monitoring of symptoms over time. Temporal segmentation of audio signals into individual words is an important pre-processing step needed prior to extracting acoustic features. Machine learning techniques could be used to automate speech segmentation via automatic speech recognition (ASR) and sequence to sequence alignment. While state-of-the-art ASR models achieve good performance on healthy speech, their performance significantly drops when evaluated on dysarthric speech. Fine-tuning ASR models on impaired speech can improve performance in dysarthric individuals, but it requires representative clinical data, which is difficult to collect and may raise privacy concerns. This study explores the feasibility of using two augmentation methods to increase ASR performance on dysarthric speech: 1) healthy individuals varying their speaking rate and loudness (as is often used in assessments of pathological speech); 2) synthetic speech with variations in speaking rate and accent (to ensure more diverse vocal representations and fairness). Experimental evaluations showed that fine-tuning a pre-trained ASR model with data from these two sources outperformed a model fine-tuned only on real clinical data and matched the performance of a model fine-tuned on the combination of real clinical data and synthetic speech. When evaluated on held-out acoustic data from 24 individuals with various neurological diseases, the best performing model achieved an average word error rate of 5.7% and a mean correct count accuracy of 94.4%. In segmenting the data into individual words, a mean intersection-over-union of 89.2% was obtained against manual parsing (ground truth). It can be concluded that emulated and synthetic augmentations can significantly reduce the need for real clinical data of dysarthric speech when fine-tuning ASR models and, in turn, for speech segmentation.

Enhancing Armenian Automatic Speech Recognition Performance: A Comprehensive Strategy for Speed, Accuracy, and Linguistic Refinement

This research introduces a comprehensive strategy to enhance the performance of an existing automatic speech recognition (ASR) model, which has been previously documented in published articles. The study sets out to achieve several objectives. Firstly, it concentrates on updating the ASR model by retraining it with new datasets. This involves integrating samples from the latest Common Voice corpus release and data collected independently via the armspeech.com web application. Another key focus lies in optimizing the ASR model for near-real-time processing, intending to improve its speed and efficiency. The proposed adjustments to the model’s architecture aim to balance accuracy and processing speed, which is essential for applications requiring prompt speech recognition. Furthermore, the research explores the integration of Transformer models into the post-processing pipeline to introduce punctuation and capitalization into the ASR output. This step not only enhances the linguistic quality of transcriptions but also improves their readability and usability. In tandem with these advancements, the research presents a systematic approach to gathering, annotating, and storing datasets specifically tailored for punctuation and capitalization tasks. The methodology outlines the acquisition and organization of a dataset conducive to training Transformer models for these linguistic tasks. This comprehensive approach, which encompasses dataset enrichment, architectural modifications, and post-processing enhancements, aims to elevate the ASR model’s accuracy, speed, and linguistic refinement, with a particular focus on addressing the intricacies of the Armenian language. The research contributes valuable insights into the optimization of ASR systems, tackling both language-specific challenges and broader issues related to linguistic post-processing.

AfriSpeech-200: Pan-African Accented Speech Dataset for Clinical and General Domain ASR

Abstract Africa has a very poor doctor-to-patient ratio. At very busy clinics, doctors could see 30+ patients per day—a heavy patient burden compared with developed countries—but productivity tools such as clinical automatic speech recognition (ASR) are lacking for these overworked clinicians. However, clinical ASR is mature, even ubiquitous, in developed nations, and clinician-reported performance of commercial clinical ASR systems is generally satisfactory. Furthermore, the recent performance of general domain ASR is approaching human accuracy. However, several gaps exist. Several publications have highlighted racial bias with speech-to-text algorithms and performance on minority accents lags significantly. To our knowledge, there is no publicly available research or benchmark on accented African clinical ASR, and speech data is non-existent for the majority of African accents. We release AfriSpeech, 200hrs of Pan-African English speech, 67,577 clips from 2,463 unique speakers across 120 indigenous accents from 13 countries for clinical and general domain ASR, a benchmark test set, with publicly available pre-trained models with SOTA performance on the AfriSpeech benchmark.

Speech corpus for Medina dialect

Automatic Speech Recognition (ASR) has standard rules which must be followed and considered carefully. Some difficulties that lead to less ASR performance is variations in pronunciation and small words misrecognition. Arabic ASR faces some challenges like difficulty in obtaining corpora for spoken dialects. Obtaining a wide range of diacritized text as well as the enormous number of word forms is considered a major challenge due to the Arabic language morphology richness and its’ letters capability to be written without diacritics. Although Arabic is one of the most popular languages, Arabic ASR systems are still rare compared with other languages. As ASR systems depend primarily on speech corpuses, Arabic ASR systems requires specific-dialect speech corpuses. Such speech corpuses are still deficient, costly, nor sometimes exists. In this research, we contribute to overcome the lack of speech recognition and misunderstanding for one of the most famous dialects in Saudi Arabia, Medina. We created a brand-new corpus “Haneen Corpus“, which consists of 70,364 tokens that have been uttered using Medina dialect, and constructed a dictionary using 64 phonemes to analyse the correct pronunciation of words. Our Medina Dialect ASR System exploited Hidden Markov Models (HMM) that achieved 92.09 % speech recognition accuracy.

Improving speech recognition systems for the morphologically complex Malayalam language using subword tokens for language modeling

This article presents the research work on improving speech recognition systems for the morphologically complex Malayalam language using subword tokens for language modeling. The speech recognition system is built using a deep neural network–hidden Markov model (DNN-HMM)-based automatic speech recognition (ASR). We propose a novel method, syllable-byte pair encoding (S-BPE), that combines linguistically informed syllable tokenization with the data-driven tokenization method of byte pair encoding (BPE). The proposed method ensures words are always segmented at valid pronunciation boundaries. On a text corpus that has been divided into tokens using the proposed method, we construct statistical n-gram language models and assess the modeling effectiveness in terms of both information-theoretic and corpus linguistic metrics. A comparative study of the proposed method with other data-driven (BPE, Morfessor, and Unigram), linguistic (Syllable), and baseline (Word) tokenization algorithms is also presented. Pronunciation lexicons of subword tokenized units are built with pronunciation described as graphemes. We develop ASR systems employing the subword tokenized language models and pronunciation lexicons. The resulting ASR models are comprehensively evaluated to answer the research questions regarding the impact of subword tokenization algorithms on language modeling complexity and on ASR performance. Our study highlights the strong performance of the hybrid S-BPE tokens, achieving a notable 10.6% word error rate (WER), which represents a substantial 16.8% improvement over the baseline word-level ASR system. The ablation study has revealed that the performance of S-BPE segmentation, which initially underperformed compared to syllable tokens with lower amounts of textual data for language modeling, exhibited steady improvement with the increase in LM training data. The extensive ablation study indicates that there is a limited advantage in raising the n-gram order of the language model beyond n=3. Such an increase results in considerable model size growth without significant improvements in WER. The implementation of the algorithm and all associated experiments are available under an open license, allowing for reproduction, adaptation, and reuse.

Using HIPAA (Health Insurance Portability and Accountability Act)-Compliant Transcription Services for Virtual Psychiatric Interviews: Pilot Comparison Study.

Automatic speech recognition (ASR) technology is increasingly being used for transcription in clinical contexts. Although there are numerous transcription services using ASR, few studies have compared the word error rate (WER) between different transcription services among different diagnostic groups in a mental health setting. There has also been little research into the types of words ASR transcriptions mistakenly generate or omit. This study compared the WER of 3 ASR transcription services (Amazon Transcribe [Amazon.com, Inc], Zoom-Otter AI [Zoom Video Communications, Inc], and Whisper [OpenAI Inc]) in interviews across 2 different clinical categories (controls and participants experiencing a variety of mental health conditions). These ASR transcription services were also compared with a commercial human transcription service, Rev (Rev.Com, Inc). Words that were either included or excluded by the error in the transcripts were systematically analyzed by their Linguistic Inquiry and Word Count categories. Participants completed a 1-time research psychiatric interview, which was recorded on a secure server. Transcriptions created by the research team were used as the gold standard from which WER was calculated. The interviewees were categorized into either the control group (n=18) or the mental health condition group (n=47) using the Mini-International Neuropsychiatric Interview. The total sample included 65 participants. Brunner-Munzel tests were used for comparing independent sets, such as the diagnostic groupings, and Wilcoxon signed rank tests were used for correlated samples when comparing the total sample between different transcription services. There were significant differences between each ASR transcription service's WER (P<.001). Amazon Transcribe's output exhibited significantly lower WERs compared with the Zoom-Otter AI's and Whisper's ASR. ASR performances did not significantly differ across the 2 different clinical categories within each service (P>.05). A comparison between the human transcription service output from Rev and the best-performing ASR (Amazon Transcribe) demonstrated a significant difference (P<.001), with Rev having a slightly lower median WER (7.6%, IQR 5.4%-11.35 vs 8.9%, IQR 6.9%-11.6%). Heat maps and spider plots were used to visualize the most common errors in Linguistic Inquiry and Word Count categories, which were found to be within 3 overarching categories: Conversation, Cognition, and Function. Overall, consistent with previous literature, our results suggest that the WER between manual and automated transcription services may be narrowing as ASR services advance. These advances, coupled with decreased cost and time in receiving transcriptions, may make ASR transcriptions a more viable option within health care settings. However, more research is required to determine if errors in specific types of words impact the analysis and usability of these transcriptions, particularly for specific applications and in a variety of populations in terms of clinical diagnosis, literacy level, accent, and cultural origin.

Disordered speech recognition considering low resources and abnormal articulation

The success of automatic speech recognition (ASR) benefits a great number of healthy people, but not people with disorders. The speech disordered may truly need support from technology, while they actually gain little. The difficulties of disordered ASR arise from the limited availability of data and the abnormal nature of speech, e.g, unclear, unstable, and incorrect pronunciations. To realize the ASR of disordered speech, this study addresses the problems of disordered speech in two respects, low resources, and articulatory abnormality. In order to solve the problem of low resources, this study proposes staged knowledge distillation (KD), which provides different references to the student models according to their mastery of knowledge, so as to avoid feature overfitting. To tackle the articulatory abnormalities in dysarthria, we propose an intended phonological perception method (IPPM) by applying the motor theory of speech perception to ASR, in which pieces of intended phonological features are estimated and provided to ASR. And further, we solve the challenges of disordered ASR by combining the staged KD and the IPPM. TORGO database and UASEECH corpus are two commonly used datasets of dysarthria which is the main cause of speech disorders. Experiments on the two datasets validated the effectiveness of the proposed methods. Compared with the baseline, the proposed method achieves 35.14%∼38.12% relative phoneme error rate reductions (PERRs) for speakers with varying degrees of dysarthria on the TORGO database and relative 8.17%∼13.00% PERRs on the UASPEECH corpus. The experiments demonstrated that addressing disordered speech from both low resources and speech abnormality is an effective way to solve the problems, and the proposed methods significantly improved the performance of ASR for disordered speech.

Türkçe Otomatik Konuşma Tanıma Sistemi için Dil Modeli Optimizasyon Yöntemi

The current Automatic Speech Recognition (ASR) modeling strategy still suffers from huge performance degradation when faced with languages with limited resources such as Turkish. Especially when the Language Model (LM) does not support the Acoustic Model (AM) sufficiently, the Word Error Rate (WER) increases. Therefore, a robust LM makes a strong contribution to improving ASR performance by generating word relations from the existing corpus. However, developing a robust language model is a challenging task due to the agglutinative nature of Turkish. Therefore, within the scope of the study, a sentence-level LM optimization method is proposed to improve the WER performance of Turkish ASR. In the proposed method, instead of a fixed word sequence obtained from the Markov assumptions, the probability of the word sequence forming a sentence was calculated. A method with n-gram and skip-gram properties is presented to obtain the word sequence probability. The proposed method has been tested on both statistical and Artificial Neural Network (ANN) based LMs. In the experiments carried out using, not only words but also sub-word level, two Turkish corpora (METU and Bogazici) shared via Linguistic Data Consortium (LDC) and a separate corpus, which we separate corpus that we specially created as HS was used. According to the experimental results obtained from statistical-based LM, 0.5% WER increases for the METU corpus, 1.6% WER decreases for the Bogazici corpus, and a 2.5% WER decrease for the HS corpus were observed. In the Feedforward Neural Networks (FNN) based LM, WER decreases were observed 0.2% for the METU corpus, 0.8% for the Bogazici corpus, and 1.6% for the HS corpus. Also, in the Recurrent Neural Network (RNN)-Long Short Term Memory (LSTM) based LM, WER decreases were observed 0.6% for METU corpus, 1.1% for the Bogazici corpus and 1.5% for the HS corpus. As a result, when the proposed method was applied to the LMs required for ASR, WER decreased, and the total performance of ASR increased.

CAST: Context-association architecture with simulated long-utterance training for mandarin speech recognition

End-to-end (E2E) models are widely used because they significantly improve the performance of automatic speech recognition (ASR). However, based on the limitations of existing hardware computing devices, previous studies mainly focus on short utterances. Typically, utterances used for ASR training do not last much longer than 15 s, and therefore the models often fail to generalize to longer utterances at inference time. To address the challenge of long-form speech recognition, we propose a novel Context-Association Architecture with Simulated Long-utterance Training (CAST), which consists of a Context-Association RNN-Transducer (CARNN-T) and a simulating long utterance training (SLUT) strategy. The CARNN-T obtains the sentence-level contextual information by paying attention to the cross-sentence historical utterances and adds it in the inference stage, which improves the robustness of long-form speech recognition. The SLUT strategy simulates long-form audio training by updating the recursive state, which can alleviate the length mismatch between training and testing utterances. Experiments on the test of the Aishell-1 and aidatatang_200zh synthetic corpora show that our model has the best recognition performer on long utterances with the character error rate (CER) of 12.0%/12.6%, respectively.

Boosting Low-Resource Speech Recognition in Air Traffic Communication via Pretrained Feature Aggregation and Multi-Task Learning

Developing a robust Automatic Speech Recognition (ASR) system usually requires a large amount of well-annotated samples which is extremely hard to build in the Air Traffic Control (ATC) due to domain-specific knowledge. In this work, we present a novel approach to improve ASR performance in the ATC domain by integrating Self-Supervised Learning and Multi-Task learning into a unified framework. Specifically, the proposed framework follows a two-stage training paradigm, i.e., (a) learning universal acoustic representations by employing the wav2vec 2.0 model and (b) jointly finetuning the model by the ASR, Speaker Role Identification, and Language Identification tasks. To capture the task-specific representations, an attention-guided feature aggregation module is dedicatedly designed to disentangle the discriminative representations from the pretrained features. In addition, the uncertainty-based loss combination strategy is employed to balance the loss weights for each task in a learnable manner. Finally, we conduct experiments to validate the technical improvements in a real-world ATC dataset. Experimental results demonstrated that the proposed framework outperforms competitive baselines among all tasks.

Two-stage deep spectrum fusion for noise-robust end-to-end speech recognition

Recently, speech enhancement (SE) methods have achieved quite good performances. However, because of the speech distortion problem, the enhanced speech may lose significant information, which degrades the performance of automatic speech recognition (ASR). To address this problem, this paper proposes a two-stage deep spectrum fusion with the joint training framework for noise-robust end-to-end (E2E) ASR. It consists of a masking and mapping fusion (MMF) and a gated recurrent fusion (GRF). The MMF is used as the first stage and focuses on SE, which explores the complementarity of the enhancement methods of masking-based and mapping-based to alleviate the problem of speech distortion. The GRF is used as the second stage and aims to further retrieve the lost information by fusing the enhanced speech of MMF and the original input. We conduct extensive experiments on an open Mandarin speech corpus AISHELL-1 with two noise datasets named 100 Nonspeech and NOISEX-92. Experimental results indicate that our proposed method significantly improves the performance and the character error rate (CER) is relatively reduced by 17.36% compared with the conventional joint training method.

Crossing language identification: Multilingual ASR framework based on semantic dataset creation & Wav2Vec 2.0

This study proposes an innovative methodology to enhance the performance of multilingual Automatic Speech Recognition (ASR) systems by capitalizing on the high semantic similarity between sentences across different languages and eliminating the requirement for Language Identification (LID). To achieve this, special bilingual datasets were created from the Mozzila Common Voices datasets in Spanish, Russian, and Portuguese. The process involves computing sentence embeddings using Language-agnostic BERT and selecting sentence pairs based on high and low cosine similarity. Subsequently, we train the Wav2vec 2.0 XLSR53 model on these datasets and assess its performance utilizing Character Error Rate (CER) and Word Error Rate (WER) metrics. The experimental results indicate that models trained on high-similarity samples consistently surpass their low-similarity counterparts, emphasizing the significance of high semantic similarity data selection for precise and dependable ASR performance. Furthermore, the elimination of LID contributes to a simplified system with reduced computational costs and the capacity for real-time text output. The findings of this research offer valuable insights for the development of more efficient and accurate multilingual ASR systems, particularly in real-time and on-device applications.

Wav2vec‐MoE: An unsupervised pre‐training and adaptation method for multi‐accent ASR

AbstractIn real life, either the subjective factors of speakers or the objective environment degrades the performance of automatic speech recognition (ASR). This study focuses on one of the subjective factors, accented speech, and attempts to build a multi‐accent ASR system to solve the degradation caused by different accents, one of whose characteristic is the low resource. To deal with the challenge of the low‐resource data and the different speech styles, a wav2vec‐MoE (mixture of experts) is proposed to adapt the wav2vec 2.0 for multi‐accent ASR. In the wav2vec‐MoE, a domain MoE is developed by introducing pseudo‐domain information in the pre‐training stage, where the domain denotes a collection of speech varied by the same influence factors. The MoE is trained with two strategies according to the proposed domain mismatch assessment between unlabeled speech and target speech, without requiring any explicit domain information. Experiments show that the wav2vec‐MoE achieves a 14.69% relative word error rate reduction (WERR) on the AESRC2020 accent dataset and an 8.79% relative WERR on the Common Voice English dataset.

Lexical modeling for the development of Amharic automatic speech recognition systems

Amharic is the second most spoken Semitic language after Arabic. It has its own syllabary writing system, each character representing a consonant and a vowel. Automatic Speech Recognition (ASR) researches for Amharic have been conducted on the basis of grapheme-based pronunciation lexicon, taking advantage of the nature of its writing system. However, the epenthetic vowel and the glottal stop consonant represented in the writing system may not be pronounced in all of their occurrences. Moreover, the writing system does not differentiate geminated and non-geminated forms of consonants. Therefore, the grapheme-based pronunciation lexicon used so far has limitations with regard to these language features. To handle these limitations, we have prepared word- and morpheme-based pronunciation lexicons using data-driven and knowledge-driven experts’ transcription. The data-driven transcription has been used for the preparation of training pronunciation lexicon while the knowledge-driven has been used to prepare morpheme- and word-based pronunciation lexicons for decoding. When morpheme-based knowledge-driven lexicons are used, better ASR performance (compared with the baseline ASR system that used grapheme-based lexicon) has been achieved although the number of phones is much more (60) than the number of phones used in the grapheme-based lexicon (37).

A Review on Grapheme-to-Phoneme Modelling Techniques to Transcribe Pronunciation Variants for Under-Resourced Language

A pronunciation dictionary (PD) is one of the components in an Automatic Speech Recognition (ASR) system, a system that is used to convert speech to text. The dictionary consists of word-phoneme pairs that map sound units to phonetic units for modelling and predictions. Research has shown that words can be transcribed to phoneme sequences using grapheme-to-phoneme (G2P) models, which could expedite building PDs. The G2P models can be developed by training seed PD data using statistical approaches requiring large amounts of data. Consequently, building PD for under-resourced languages is a great challenge due to poor grapheme and phoneme systems in these languages. Moreover, some PDs must include pronunciation variants, including regional accents that native speakers practice. For example, recent work on a pronunciation dictionary for an ASR in Iban, an under-resourced language from Malaysia, was built through a bootstrapping G2P method. However, the current Iban pronunciation dictionary has yet to include pronunciation variants that the Ibans practice. Researchers have done recent studies on Iban pronunciation variants, but no computational methods for generating the variants are available yet. Thus, this paper reviews G2P algorithms and processes we would use to develop pronunciation variants automatically. Specifically, we discuss data-driven techniques such as CRF, JSM, and JMM. These methods were used to build PDs for Thai, Arabic, Tunisian, and Swiss-German languages. Moreover, this paper also highlights the importance of pronunciation variants and how they can affect ASR performance.

Finnish parliament ASR corpus: Analysis, benchmarks and statistics.

Public sources like parliament meeting recordings and transcripts provide ever-growing material for the training and evaluation of automatic speech recognition (ASR) systems. In this paper, we publish and analyse the Finnish Parliament ASR Corpus, the most extensive publicly available collection of manually transcribed speech data for Finnish with over 3000h of speech and 449 speakers for which it provides rich demographic metadata. This corpus builds on earlier initial work, and as a result the corpus has a natural split into two training subsets from two periods of time. Similarly, there are two official, corrected test sets covering different times, setting an ASR task with longitudinal distribution-shift characteristics. An official development set is also provided. We developed a complete Kaldi-based data preparation pipeline and ASR recipes for hidden Markov models (HMM), hybrid deep neural networks (HMM-DNN), and attention-based encoder-decoders (AED). For HMM-DNN systems, we provide results with time-delay neural networks (TDNN) as well as state-of-the-art wav2vec 2.0 pretrained acoustic models. We set benchmarks on the official test sets and multiple other recently used test sets. Both temporal corpus subsets are already large, and we observe that beyond their scale, HMM-TDNN ASR performance on the official test sets has reached a plateau. In contrast, other domains and larger wav2vec 2.0 models benefit from added data. The HMM-DNN and AED approaches are compared in a carefully matched equal data setting, with the HMM-DNN system consistently performing better. Finally, the variation of the ASR accuracy is compared between the speaker categories available in the parliament metadata to detect potential biases based on factors such as gender, age, and education.