I’m noticing that the gpt-4o-transcribe model often drops the end of recordings.
It’s not that it truncates the response (like this post) to some time limit, it’s that it ignores the last phrase or sentence, even in very short recordings.
Here’s the output of three models for the same input.
gpt-4o-transcribe: Testing pineapples and cats, as well as cheetahs.
gpt-4o-mini-transcribe: Testing pineapples and cats, as well as cheetahs, which are cats.
DeepGram: Testing pineapples and cats as well as cheetahs, which are cats.
Note that these recordings are quite bad quality and that this happens about 1 in every 8 recordings I try. Most of my recordings are just a few sentences.
Another example:
gpt-4o-transcribe: This gives me a single byte string, doesn’t it?
gpt-4o-mini-transcribe: This gives me a single byte string, doesn’t it? I thought tobytes gave me an array of bytes, but not by version.
DeepGram: This gives me a single byte string, doesn’t it? I thought two bytes gave me an array of bytes, the NumPy version.