KoelLabs/speech-accent-context

2138 English speakers from a variety of language backgrounds (pulled from the Speech Accent Archive) saying the sentence

"Please call Stella. Ask her to bring these things with her from the store: Six spoons of fresh snow peas, five thick slabs of blue cheese, and maybe a snack for her brother Bob. We also need a small plastic snake and a big toy frog for the kids. She can scoop these things into three red bags, and we will go meet her Wednesday at the train station."

We focus on three words -- "snack", "snake", "bags" -- which are notoriously tricky for humans to identify out-of-context. We use an automated script to transcribe them both in-context and clipped with no-context using the Facebook 60 phoneme transcription model. For cases of mispronunciation, the results show that in-context, the transformer layers of the Wav2Vec2 model act as a language model to hallucinate the more "standard" g2p transcription of the words whereas out-of-context, the model predicts a transcription closer to what the speaker actually said.