[Paper Review] OkwuGbé: End-to-End Speech Recognition for Fon and Igbo

Bonaventure F. P. Dossou, Chris Chinenye Emezue|arXiv (Cornell University)|Mar 13, 2021

Speech Recognition and Synthesis51 references4 citations

TL;DR

This paper proposes end-to-end speech recognition models for Fon and Igbo using Word2Vec and Poincaré word embeddings, demonstrating that these embeddings capture semantic and hierarchical relationships in African languages. Key results include 60–62% transfer learning accuracy across languages and evidence of bias mitigation through fine-tuning, advancing NLP for low-resource African languages.

ABSTRACT

Language is inherent and compulsory for human communication. Whether expressed in a written or spoken way, it ensures understanding between people of the same and different regions. With the growing awareness and effort to include more low-resourced languages in NLP research, African languages have recently been a major subject of research in machine translation, and other text-based areas of NLP. However, there is still very little comparable research in speech recognition for African languages. Interestingly, some of the unique properties of African languages affecting NLP, like their diacritical and tonal complexities, have a major root in their speech, suggesting that careful speech interpretation could provide more intuition on how to deal with the linguistic complexities of African languages for text-based NLP. OkwuGbé is a step towards building speech recognition systems for African low-resourced languages. Using Fon and Igbo as our case study, we conduct a comprehensive linguistic analysis of each language and describe the creation of end-to-end, deep neural network-based speech recognition models for both languages. We present a state-of-art ASR model for Fon, as well as benchmark ASR model results for Igbo. Our linguistic analyses (for Fon and Igbo) provide valuable insights and guidance into the creation of speech recognition models for other African low-resourced languages, as well as guide future NLP research for Fon and Igbo. The Fon and Igbo models source code have been made publicly available.

Motivation & Objective

• To develop word embedding models tailored for African Indigenous Languages (AILs), specifically Fon and Nobiin, which lack dedicated embeddings.
• To investigate whether Poincaré embeddings can effectively model hierarchical linguistic structures in low-resource languages.
• To evaluate transfer learning performance between Fon and Nobiin embeddings for named entity recognition tasks.
• To explore linguistic and social biases in word embeddings and propose mitigation strategies.
• To promote the creation of language-specific NLP resources for African languages to improve downstream task performance.

Proposed method

• Trained Word2Vec models on large monolingual corpora of Fon and Nobiin, using continuous skip-gram architecture with negative sampling.
• Applied Poincaré embeddings to model hierarchical relationships in word representations, leveraging hyperbolic geometry for tree-structured data.
• Optimized hyperparameters including dimension size (d=10 for Nobiin, d=50 for Fon) and negative curvature (c=15, 20) for Poincaré models.
• Evaluated transfer learning by testing Fon Poincaré model on Nobiin names and vice versa, using classification reports and global accuracy.
• Mitigated gender bias in embeddings by fine-tuning the maximum distance between predicted and target words in the context window.
• Visualized word clusters to analyze semantic groupings and bias patterns, particularly around gendered roles and housekeeping terms.

Experimental results

Research questions

• RQ1 Can Poincaré embeddings effectively model hierarchical and semantic relationships in low-resource African languages like Fon and Nobiin?
• RQ2 To what extent can pre-trained word embeddings from one African language be transferred to improve named entity recognition in another?
• RQ3 How do existing societal biases manifest in word embeddings trained on African language corpora, and can they be mitigated?
• RQ4 What is the performance of Word2Vec and Poincaré embeddings in capturing semantic similarity and syntactic relationships in Fon and Nobiin?
• RQ5 Can transfer learning between Fon and Nobiin embeddings yield meaningful improvements in downstream NLP tasks despite linguistic differences?

Key findings

• The Fon Poincaré model achieved 60% global accuracy when tested on the Nobiin named entity recognition dataset, indicating effective cross-lingual transfer.
• The Nobiin Poincaré model achieved 62% global accuracy when tested on the Fon named entity recognition dataset, demonstrating bidirectional transfer capability.
• Poincaré embeddings with negative curvature c=20 and dimension d=10 showed the best performance on Nobiin, achieving 50% global accuracy in classification.
• Word2Vec models revealed gender bias, clustering housekeeping terms (e.g., floor-sweeping, washing) closely with 'mother', reflecting societal stereotypes.
• Fine-tuning the maximum distance in the context window reduced bias, improving clustering of 'mother' with 'father', 'sister', and 'brother'.
• The study confirms that word embeddings trained on African language corpora reflect existing societal biases, emphasizing the need for bias mitigation in NLP models.