It help to do the math right.

app/src/main/java/com/mjm/whisperVoiceRecognition/VoiceKeyboardInputMethodService.java

@@ -123,7 +123,7 @@ public boolean onTouch(View view, MotionEvent motionEvent) {
                 Pair<Optional<float[]>, Long> byteBuffer = RustLib.endRec();
 
                 if (byteBuffer.first.isPresent()) {
-//                    draw(byteBuffer.first.get());
+                    draw(byteBuffer.first.get());
 //                    signalTxt.setText(new String(byteBuffer.second / 1000_000 + " ms"));
                     Pair<String, Long> transcribeAudio = transcribeAudio(byteBuffer.first.get());
                     String transcribed = transcribeAudio.first.trim() + " ";
@@ -210,7 +210,7 @@ private void draw(float[] floats) {
         }
         //write the bitmap to file
         try {
-            FileOutputStream out = new FileOutputStream(getFilesDir().getAbsolutePath() + "/spectrogram3.png");
+            FileOutputStream out = new FileOutputStream(getFilesDir().getAbsolutePath() + "/spectrogram" + System.nanoTime() + ".png");
             bitmap.compress(Bitmap.CompressFormat.PNG, 100, out);
             out.close();
         } catch (IOException e) {

app/src/main/rust/src/lina.rs

@@ -32,7 +32,7 @@ pub(crate) fn dot_product(left: &[f32], right: &[f32]) -> f32 {
 }
 #[cfg(target_arch = "aarch64")]
 // #[cfg(target_arch = "aarch64")]
-/// I tried to beat the compiler. I lost. Badly. I suspect the compiler detects both the loop above and the looped function calls correctly and does a proper matrix multiplication
+/// I tried to beat the compiler. I lost. Badly. This is because the latency cost for the vector instruction outweighs the benefit to using them, UNLESS, we allow the compiler to hide the latency behind many such calls at once:loop unrolling.
 /// Calculates the dot product of two slices of `f32` values.
 ///
 /// # Parameters

app/src/main/rust/src/spectrogram.rs

@@ -120,16 +120,18 @@ fn compute_power(fft_work_buffer: &[Complex32], n_fft: usize) -> Vec<f32> {
     power_spectrum[0..n_fft].to_vec()
 }
 
-/// Append the log mel spectrogram to the mel spectrogram columns buffer.
-fn append(table: &mut Vec<Vec<f32>>, column: &mut Vec<f32>) {}
-
 /// Normalize the mel spectrogram columns buffer.
 fn normalize(mel_spectrogram_columns: &mut [f32]) {
     // Compute the maximum value of the mel spectrogram columns buffer.
+    mel_spectrogram_columns
+        .iter_mut()
+        .for_each(|x| *x = (*x).max(1e-10).log10());
+
     let maximum_value = mel_spectrogram_columns
         .iter()
         .fold(-1e20f32, |acc, f| f.max(acc));
+
     mel_spectrogram_columns
         .iter_mut()
-        .for_each(|x| *x = ((*x).max(1e-10).log10().max(maximum_value - 8.0) + 4.0) / 4.0);
+        .for_each(|x| *x = ((*x).max(maximum_value - 8.0) + 4.0) / 4.0);
 }