This paper describes new reference benchmark results based on hybrid Hidden Markov Model and Deep Neural Networks (HMM-DNN) for the GlobalPhone (GP) multilingual text and speech database. GP is a multilingual database of high-quality read speech with corresponding transcriptions and pronunciation dictionaries in more than 20 languages. Moreover, we provide new results for five additional languages, namely, Amharic, Oromo, Tigrigna, Wolaytta, and Uyghur. Across the 22 languages considered, the hybrid HMM-DNN models outperform the HMM-GMM based models regardless of the size of the training speech used. Overall, we achieved relative improvements that range from 7.14% to 59.43%.

This paper presents our investigation of the effect of language resources on the performance of Amharic speech recognition. We have used language model training text of different sizes and seen the effect on word error rate (WER) reduction. Moreover, we have investigated the effect of handling language issues (germination, epenthetic vowel insertion and glottal stop consonant pronunciation) on the performance of speech recognition systems using data-driven phone-level transcriptions. The results of our experiments show that only slight reduction in WER can be obtained by increasing language model training text. However, proper transcription of gemination, the epenthetic vowel and the glottal stop consonant did not bring performance improvement for Amharic speech recognition. This can be attributed to the larger number of phone HMM acoustic models (62 compared to 37 phone set of the grapheme-based phone-level transcriptions) trained with a small (5 hrs) training speech.

More than 7000 languages are spoken in the world today. Amharic is one of the languages spoken in the East African country Ethiopia. A lot of speech data is being made every day in different languages as machines are getting better at processing and have improved storing capacity. However, searching for a particular word with its respective time frame inside a given audio file is a challenge. Since Amharic has its own distinguishing characteristics, such as glottal, palatal, and labialized consonants, it is not possible to directly use models that are developed for other languages. A popular approach in developing systems for searching particular information in speech involves using an automatic speech recognition (ASR) module that generates the text version of the speech where the word or phrase is searched based on text query. However, it is not possible to transcribe a long audio file without segmentation, which in turn affects the performance of the ASR module. In this paper, we are reporting our investigation on the effects of manual and automatic speech segmentation of Amharic audio files in a spiritual domain. We have used manual segmentation as a baseline for our investigation and found out that sentence-like automatic segmentation resulted in a word error rate (WER) closer to the WER achieved on the manually segmented test speech. Based on the experimental results, we propose Amharic speech search using text word query (ASSTWQ) based on automatic sentence-like segmentation. Since we have achieved lower WER using the previously developed speech corpus, which is in a broadcast news domain, together with the in-domain speech corpus, we recommend using both in- and out-domain speech corpora to develop the Amharic ASR module. The performance of the proposed ASR is a WER of 53% that needs further improvement. Combining two language models (LMs) developed using training text from the two different domains (spiritual and broadcast news) allowed a WER reduction from 53% to 46%. Therefore, we have developed two ASSTWQ systems using the two ASR modules with WERs of 53% and 46%.

Nowadays, social media platforms provide space that allows communication and sharing of various resources using a variety of natural languages in different cultural and multilingual aspects. Although this interconnectedness offers numerous benefits, it also exposes users to the risk of encountering offensive (OFFN) and harmful content, including hateful speech. In order to create a model for detecting hateful content in resource‐rich languages, lexicons, word embedding, topic modeling, and transformer language models were applied. Low‐resource languages, including Ethiopian languages, suffering in lack of such linguistic resources. Multilingual hateful content detection brings complex challenges due to cultural and linguistic varieties. The paper proposes a multilingual hateful content identification model using a transformer language model and hybrid lexicon techniques to enhance hateful content recognition in low‐resource Ethiopian languages. First, hateful content disseminated on Facebook in Ethiopian‐languages target was identified as (insult, identity hate, antagonistic, and threat) using topic modeling techniques. Then, we compiled different hateful terms from sources such as guidelines and proclamations related to the Ethiopian context. We created Ethiopian context‐based transformer language models. We utilized topic words‐based datasets to construct pretrained transformer language models and multilingual lexicons of major Ethiopian languages. Finally, their performance was compared by integrating them into deep learning‐based low‐resource Ethiopian languages’ hateful content detection framework. Among applied deep learning algorithms with Ethiopian language linguistic resources, word2vec‐based multilingual lexicons with convolutional neural network (CNN) outperform than others. The result indicated that constructing topic words based multilingual word2vec lexicons outperformed than transformers language model based on topics modeling for low‐resource Ethiopian languages, effectively produce the promising hate speech (HATE) detection approach of low‐resource Ethiopian languages.