🎤 FonBench Leaderboard

ASR Benchmark and Adaptation for the Fon language

A low-resource African tonal language (~8M speakers in Benin)

Test set: 3,541 examples · 7 speakers · stratified by speaker

Six multilingual ASR models evaluated zero-shot on the Fon test set

Filter by model family

Encoder-CTC (SSL) Encoder-only (SSL) Encoder-Decoder

Show disqualified models (CER > 200%)

Model	Family	Params (M)	WER (%)	CER (%)	T-WER (%)	WER_seg (%)	WER_ton (%)	RTFx	Score	Status
Whisper-small	Encoder-only (SSL)	2300	101.03	360.12	152.41	100.95	46.43	111.81	0.4962	❌ Disqualified (CER>200%)

Plot

Note on model selection: Although XLSR-53 obtains the highest composite score, its WER_segmental exceeds 100% (no exploitable transcription). MMS-1b is retained for Phase 2 based on WER and CER (the only diacritic-agnostic metrics in this corpus, whose reference transcriptions do not include systematic tonal marking). See thesis Section 4.1.2 for the full justification.