whisper_full_get_token_text crash by run in a loop in swift

[bitsmakerde]

Hi,

I have the problem that the translation with files which are bigger as 30 seconds are terrible, in the swiftUI example.

to solve this problem I want split my files in 30 seconds parts and loop over them to get the best result with init prompts too.

But if I have more as one file my app crash and I have no idea to fix it.

I get this error:

Selecting 8 threads
About to run whisper_full
WHISPER_ASSERT: /Users/andre/Documents/GitHub/whisper.cpp/whisper.cpp:4531: n_logits == ctx.vocab.n_vocab

here how I loop the my files:

func transcribeSample() {
        if let sampleUrl {
            let segments = [sampleUrl, sampleUrl]

            for segment in segments {
                print("segment: \(segment)")
                transcribeAudio(segment)
            } 
        } else {
            messageLog += "Could not locate sample\n"
        }
    }

And here my config

func fullTranscribe(samples: [Float]) {
        // Leave 2 processors free (i.e. the high-efficiency cores).
        let maxThreads = max(1, min(8, cpuCount() - 2))
        print("Selecting \(maxThreads) threads")
        let myString = "ähm, ah, äh, ahm, äähm, ääh, ähh, uhm, eh, ehm, hmm, mm, mhm, mmm, uh, um. I mean, mean."
        var prompt: UnsafePointer<CChar>?

        myString.withCString { cStr in
            // cStr ist ein UnsafePointer<CChar>, der auf den C-String zeigt
            prompt = UnsafePointer<CChar>(cStr)

            // Beispielhafte Nutzung des Zeigers
        }

        var params = whisper_full_default_params(WHISPER_SAMPLING_GREEDY)
        "de".withCString { en in
            // Adapted from whisper.objc
            params.print_realtime = true
            params.print_progress = false
            params.print_timestamps = true
            params.print_special = true
            params.translate = false
            params.language = en
            params.n_threads = Int32(maxThreads)
            params.offset_ms = 0
            params.no_context = true
            params.single_segment = false
            params.suppress_blank = false
            params.initial_prompt = prompt

            params.max_len = 1
            params.split_on_word = true
            params.token_timestamps = true

            whisper_reset_timings(context)
            print("About to run whisper_full")
            samples.withUnsafeBufferPointer { samples in
                if whisper_full(context, params, samples.baseAddress, Int32(samples.count)) != 0 {
                    print("Failed to run the model")
                } else {
                    whisper_print_timings(context)
                }
            }
        }
    }

[bitsmakerde]

ok I found a little bit more out, there my app is crashing.

so the first run works perfect, but the second run it's crashing by this line of code:

String(cString: whisper_full_get_token_text(context, i, token))

I want every word together with the start and end time back. I have no idea why if runs one time but the second not. the context is the same like by the first run. the model and the sample file too.