Currently, the scoring model is based on standard American English phonemes, so it works best if that's your target accent. British or Australian pronunciations may score lower even when correct.
Supporting multiple accent references is on the roadmap — the architecture allows swapping the wav2vec2 model, so it's technically feasible.
Happy to accept contributions if anyone wants to explore this!
Supporting multiple accent references is on the roadmap — the architecture allows swapping the wav2vec2 model, so it's technically feasible. Happy to accept contributions if anyone wants to explore this!