Hi everyone,
I’ve been using the realtime transcription API via WebSocket for quite a while. My configuration follows the transcription_session.update documentation, and I always include the following setting:
"include": ["item.input_audio_transcription.logprobs"]
Up until recently, the responses consistently contained the logprobs
field, which is essential for our downstream post-processing. Here’s an example of the response we used to receive:
{
"type": "conversation.item.input_audio_transcription.completed",
"event_id": "event_BeBydw2SremYvspy4YZJa",
"item_id": "item_BeByZacSM5nU3i5e0Gidj",
"content_index": 0,
"transcript": "Xin lỗi em không hiểu.",
"logprobs": [
{ "token": "Xin", "logprob": -5.2024107, "bytes": [88, 105, 110] },
{ "token": " lỗi", "logprob": -0.7330262, "bytes": [32, 108, 225, 187, 151, 105] },
{ "token": " em", "logprob": -1.6249285, "bytes": [32, 101, 109] },
{ "token": " không", "logprob": -1.3074652, "bytes": [32, 107, 104, 195, 180, 110, 103] },
{ "token": " hiểu", "logprob": -2.357233, "bytes": [32, 104, 105, 225, 187, 131, 117] },
{ "token": ".", "logprob": -0.22520193, "bytes": [46] }
]
}
However, as of today, we’ve noticed that none of the transcription responses include the logprobs
field anymore, even though our configuration hasn’t changed.
Has there been an update to the API that removed logprobs
from responses?
If so, could someone from the team confirm this and let us know when this change was applied?
Thanks in advance!