How can I improve the quality by input part of the script?

[iTula]

We are using this app to transcribe a meeting. Some or part of the speeches in meeting are based on a script. How can I improve the quality of transcription using the script?

[Tamnac]

pass it in as the --initial_prompt (won't work with distil models afaik)

[iTula]

How can I do that? Can you give me an example? Thank you a lot. 发自我的 iPhone 在 2024年8月4日，03:49，Tamnac ***@***.***> 写道：  pass it in as the --initial_prompt (won't work with distil models afaik) — Reply to this email directly, view it on GitHub<#259 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/BF6URL336JO6YIURVE5IYC3ZPUX43AVCNFSM6AAAAABHVJPV7GVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTAMRTGIYTIOI>. You are receiving this because you authored the thread.Message ID: ***@***.***>

[Tamnac]

I assume you know how to use the cli, so you just do something like `.\faster-whisper-xxl.exe --initial_prompt "<script>".

That said, the prompt is more for like context/topic, if the script is included in the speeches word for word, you might be better of passing in the transcript + script into something like chatgpt and asking it to combine them.

[ClaireCJS]

I assume you know how to use the cli, so you just do something like `.\faster-whisper-xxl.exe --initial_prompt "<script>".

That said, the prompt is more for like context/topic, if the script is included in the speeches word for word, you might be better of passing in the transcript + script into something like chatgpt and asking it to combine them.

I'm in a slightly different situation and wondering if you could offer your advice.

A lot of music has a TXT download of lyrics, but no SRT/LRC.

The AI often guesses the wrong word when making SRC/LRC from scratch.

Is there any way to not prompt it, but prime it with the TXT file, so that it doesn't make incorrect guesses? Basically, if I have the lyrics already, I just want it to add the timing.

[Tamnac]

I think you'll have to do the second suggestion. Make a transcript using this, then give it with timestamps + your txt file to an llm like chatgpt and ask it to combine the two using the txt as the content and the transcript for timestamps

[ClaireCJS]

I think you'll have to do the second suggestion. Make a transcript using this, then give it with timestamps + your txt file to an llm like chatgpt and ask it to combine the two using the txt as the content and the transcript for timestamps

That is a shame. Could there be a way to add a TXT to it in advance, and perhaps it could then be turned into a prompt along the lines of "using these lyrics, do "?

[iTula]

Thanks, I will try that out.

…

________________________________ 发件人: Tamnac ***@***.***> 发送时间: 2024年8月6日 19:40 收件人: Purfview/whisper-standalone-win ***@***.***> 抄送: iTula ***@***.***>; Author ***@***.***> 主题: Re: [Purfview/whisper-standalone-win] How can I improve the quality by input part of the script? (Discussion #259) I assume you know how to use the cli, so you just do something like `.\faster-whisper-xxl.exe --initial_prompt "<script>". That said, the prompt is more for like context/topic, if the script is included in the speeches word for word, you might be better of passing in the transcript + script into something like chatgpt and asking it to combine them. ― Reply to this email directly, view it on GitHub<#259 (reply in thread)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/BF6URL6GC2O4S2XED5OAGGLZQCY2DAVCNFSM6AAAAABHVJPV7GVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTAMRVGI4DOMQ>. You are receiving this because you authored the thread.Message ID: ***@***.***>