Systematic evaluation of tonal fidelity in facebook/omniASR-CTC-1B when processing Igbo, a tonal Niger-Congo language with ~45 million speakers.
# Igbo ASR Tonal Evaluation
Systematic evaluation of tonal fidelity in facebook/omniASR-CTC-1B when processing Igbo, a tonal Niger-Congo language with ~45 million speakers.
## Overview
This project reveals systematic tonal diacritic loss in a state-of-the-art multilingual ASR model:
- **75.5% diacritic loss** on tonal markers (bootstrap 95% CI: [57.1%, 89.7%])
- **Minimal pair collapse**: Model cannot distinguish phonemically contrastive tones
- **Orthographic bias**: Model hallucinates tone marks on monotone speech
**Key Insight:** The model appears to generate diacritics probabilistically based on lexical priors rather than acoustic conditioning.
## Dataset
**21 audio samples** across 4 error categories:
1. Cross-lingual Orthographic Interference (5 samples)
2. Phonemic Tone Sensitivity (6 samples)
3. Language Boundary Effects (5 samples)
4. Domain-Specific Lexical Coverage (5 samples)
**View Dataset on HuggingFace**
### Listen to Examples
Audio files are included in this repository (M4A format). Click to play directly on GitHub:
**Tonal Minimal Pairs:**
- 06_tonal_akwa.m4a - 4 different words collapsed to random outputs
**Monotone Hallucination:**
- 09_tonal_flat.m4a - Flat speech, model ADDED tones that weren't spoken
**Code-Switching:**
- 11_codeswitch_en2ig.m4a - English perfect, Igbo loses tones
## Quick Start
### Installation
```bash
git clone
github.com
cd igbo-asr-tonal-evaluation
pip install -r requirements.txt
```
### Run Analysis
```bash
jupyter notebook analysis.ipynb
```
Or open in Google Colab:
## Repository Structure
```
igbo-asr-tonal-evaluation/
├── data/
│ ├── audio/
│ │ ├── 01_script_names.m4a # Cross-lingual interference samples
│ │ ├── 02_script_formal.m4a
│ │ ├── 03_script_numbers.m4a
│ │ ├── 04_script_proverb.m4a
│ │ ├── 05_script_slow.m4a
│ │ ├── 06_tonal_akwa.m4a # Tonal minimal pairs
│ │ ├── 07_tonal_oke.m4a
│ │ …