For a tonal language, a speech transcript that omits tone can erase distinctions that matter to both meaning and evaluation. This corpus makes tone part of the supervision: its transcribed utterances use Fongbe orthography with tone marks, and a second annotator reviews the transcriptions. It also preserves validated recordings that have not yet been transcribed, making the resource useful beyond conventional supervised ASR.
What Sets It Apart
- The transcribed and untranscribed subsets serve different stages of model development: the first supports supervised recognition, while the second can contribute to self-supervised or semi-supervised training.
- Recordings come from volunteer speakers using mobile phones, offering speech collected outside a strictly controlled studio setting.
- Pseudonymous speaker IDs allow researchers to identify recordings from the same voice across subsets without exposing speaker names or personal information.
- The release documents its transcription conventions, including tone marks and special Fongbe characters. Unicode normalization is important when comparing or scoring text.
Who It’s For—and the Trade-Off
A good fit for researchers building or evaluating Fongbe ASR, adapting speech encoders, or studying tone-aware speech recognition in a low-resource language. It is less suitable for projects that need broad multilingual coverage, studio-only audio, or transcripts for every recording: the corpus focuses on Fongbe, includes mobile-phone recordings, and only one subset has transcriptions.