Prompting in Whisper to INCLUDE punctuation in transcription

Prompting
1

Hi
I have been sucessfully using Whisper for some time but find it sometimes does not transcribe some puctuation (particularly ‘new paragraph’) . I want the words ‘new paragraph’ left in the transcription.

            transcript = openai.Audio.transcribe(
                model = "whisper-1",
                file= audio_file,
                prompt=(None, "Include 'new paragraph, paragraph and other punctuation. including (brackets) in the transcription"
                        ),
                options={
                    "language": "en",
                    "temperature": "0"
                    }
                )

Any thoughts on what prompt I should use - this does sometime work?

Should I raise the temperature - I don’t want Whisper to be too creative as it is transcribing medical letters!

Any thoughts gratefully received.

Justin

2

The purpose of the transcriptions prompt field is not to guide the AI model directly. Rather, it serves to offer preceding transcribed text up to the point where the audio commences. This approach aids the AI in determining the language and crafting an appropriate start to the audio file transcription.

3

Many thanks for taking the trouble to respond.
From what you write, it appears that I should write something more like what the speaker will actually say in the prompt.
EG
“here is some text, new paragraph, Here is some more text.”

Is that correct?

4

You could conceivably write something that is not just a lead-up, but is also demonstrative of the punctuation you want. You aren’t charged for this input, and no maximum length is stated.

prompt = “Here’s today’s radio show.\nWe are proud to welcome a new presenter today. He is a native English speaker.\nWelcome, everybody.”

You’ll likely need another AI to create logical paragraphs though.

5

Hi

Thanks for this advice. My aim is to actually include the text ‘new paragraph’ in the response from whisper every time the speaker says it. I have additional code that processes the text to actually produce a new paragraph.

My issue is that with the following , it sometimes does include the text and sometimes doesn’t.

transcript = openai.Audio.transcribe(
model = “whisper-1”,
file= audio_file,
prompt=(None, “Include all punctuation, new paragraph, including (brackets), new paragraph, in the transcription”
),
options={
“language”: “en”,
“temperature”: “0”
}
)

I have a attached snippet from a longer audio file that works, however when the full audio file was processed, New paragraph was ommited.

Any advice gratefully received.

Justin

PS When using microsoft’s transribe, it does transcribe new paragraph - I assume this is using Whisper?

Working audio snippet