| I'm deaf. Something close to standard Canadian English is my native language. Most native English speakers claim my speech is unmarked but I think they're being polite; it's slightly marked as unusual and some with a good ear can easily tell it's because of hearing loss. Using the accent guesser, I have a Swedish accent. Danish and Australian English follow as a close tie. It's not just the AI. Non-native speakers of English often think I have a foreign accent, too. Often they guess at English or Australian. Like I must have been born there and moved here when I was younger, right? I've also been asked if I was Scandinavian. Interestingly I've noticed that native speakers never make this mistake. They sometimes recognize that I have a speech impediment but there's something about how I talk that is recognized with confidence as a native accent. That leads me to the (probably obvious) inference that whatever it is that non-native speakers use to judge accent and competency, it is different from what native speakers use. I'm guessing in my case, phrase-length tone contour. (Which I can sort of hear, and presumably reproduce well, even if I have trouble with the consonants.) AI also really has trouble with transcribing my speech. I noticed that as early as the '90s with early speech recognition software. It was completely unusable. Even now AI transcription has much more trouble with me than with most people. Yet aside from a habit of sometimes mumbling, I'm told I speak quite clearly, by humans. Hearing different things, as it were. |
I don't know what your transcription use cases are, but you may be able to get an improvement by fine-tuning Whisper. This would require about $4 in training costs[1], and a dataset with 5-10 hours of your labeled (transcribed) speech, which may be the bigger hurdle[2].
1. 2000 steps took me 6 hours on an A100 on Collab, fine-tuning openai/whisper-large-v3 on 12 hours of data. I can shar my notebook/script with you if you'd like.
2. I am working on a PWA that makes it simple for humans to edit initial, automated transcriptions with mistakes for feeding the correct dataset back into the pipeline for fine-tuning, but its not ready yet