Emperor-Trillion/yoruba-tts-evaluation

A lab-based evaluation of the Yorùbá Text-to-Speech model using SpeechT5 # yoruba-tts-evaluation A lab-based evaluation of the Yorùbá Text-to-Speech model using SpeechT5 Title: Evaluating Yorùbá TTS Using SpeechT5: A Lab-Based Study on Model Performance and Challenges Author: Sunday Emmanuel Sanni Date: May 2025 --- Introduction Text-to-Speech (TTS) systems have rapidly advanced in the past decade, primarily for high-resource languages. However, languages like Yorùbá, spoken by over 40 million people, remain underrepresented in speech AI research. This short study investigates the quality and challenges of Yorùbá speech synthesis using a pre-trained TTS model built on Microsoft’s SpeechT5 architecture. --- Objective The goal of this lab work was to evaluate the performance of the `imhotepai/yoruba-tts` model, a Yorùbá TTS system available on Hugging Face, through direct model interaction and objective quality metrics. --- Model Overview - Base model: SpeechT5ForTextToSpeech - Architecture: Transformer encoder-decoder - Tokenizer: SentencePiece - Input: Tokenized Yorùbá text - Output: Mel-spectrograms and synthesized waveforms - Conditioning: Speaker embeddings from pretrained file - Vocoder: SpeechT5HifiGan --- Experimental Setup - Platform: Google Colab - Tools Used: Hugging Face Transformers, PyTorch, Librosa - Model repo: ImhotepAI/yoruba-tts Sample Input Text: ```python speech_list = [ "Ìfẹ́ ni ògbóni yànìyàn", "Àgbọ̀nkọ́lọ́ Olórun dáni", "Ìfẹ́ Ọlọ́run", "Olúkòso", "Olórí kòdi ọ̀rọ̀ ara rẹ̀ mú", "Ìwàlàwà ọmọ obìnrin" ] ``` - Sample rate: 16,000 Hz - Voice identity: Default speaker\_embeddings.pt --- Evaluation Metrics | Metric | Score | Interpretation | | ---------------------------------------------- | ----- | -------------------------------------- | | MCD (Mel-Cepstral Distortion) | 77.46 | High spectral distortion, poor quality | | PESQ (Perceptual Evaluation of Speech Quality) | 1.14 | Close to lowest possi …