Traditional Chinese to Simplified Chinese

[777sfdf]

Testing on the Windows platform that the text (subtitles) transcribed using Chinese audio are in Traditional Chinese. If you want to output Simplified Chinese, how should you solve it? If you have the time to reply and answer my questions, I would greatly appreciate it. Thanks for you!!!

The core code is as follows: var segments = new List(); var encoderBegins = new List();
string ModelFilePath = "ggml-base.bin"; string txtFilePath = "mqfwn-f344o.wav";

            using var factory = WhisperFactory.FromPath(ModelFilePath);

            using var processor = factory.CreateBuilder()
                            .WithLanguage("zh")
                            .WithEncoderBeginHandler((e) =>
                            {
                                encoderBegins.Add(e);
                                return true;
                            })
                            .WithSegmentEventHandler(segments.Add)
                            .Build();

            using var fileReader = File.OpenRead(txtFilePath);
            await foreach (var result in processor.ProcessAsync(fileReader))
            {
                Console.WriteLine("语音转文字转换成功: " + $"{result.Start}->{result.End}: {result.Text}");

            }

        }

[sandrohanea]

Hello, I'm not sure neither I tested if this works, but based on the original whisper discussion, prompting might help in this scenario: https://github.com/openai/whisper/discussions/277

If you can test it, and update here if that works, it would be great.

[777sfdf]

Thank you very much for your answer. I'm sorry for the delay in replying to you. I have already read this issue, but due to my lack of proficiency in. net, what puzzled me was how to use this command in a project - initial_ Prompt, so if you could have time to revisit this question and provide an answer, I would greatly appreciate it. Thank you

[sandrohanea]

In order to use the initial prompt, you will call WithPrompt method on the whisperBuilder: https://github.com/sandrohanea/whisper.net/blob/454ad43043e3b5cd920e5e3a1cb309861c21d158/Whisper.net/WhisperProcessorBuilder.cs#L286C36-L286C46

[777sfdf]

Okay, thank you again for your prompt response. I will provide feedback on the results after the test

[sandrohanea]

Hey @777sfdf ,

Any news about the prompt effectiveness for your use-case?

[777sfdf]

I'm very sorry for not replying to your reply in a timely manner. The main reason is that during the testing a while ago, the results were not very good. Today, I conducted the testing again and finally achieved good results. The audio escaped text is already Simplified Chinese

The following code has been added with only one additional method, WithPrompt. The rest, including the model and audio, have not changed

The rendering is as follows

In addition, I would also like to inquire about the accuracy ranking of the four models included in the ggml model. Is it in the order of ting<base<small<medium? I hope to receive your reply. Thank you

[sandrohanea]

Awesome, I'm glat to here that "WithPrompt" is working as expected in the given scenario.

Also, as a suggestion, if you already know the language, instead of auto you can use zh so that detection of the language won't take place and you'll get transcripts faster.

About the model sizes, they are: tiny < base < small < medium < large