The provided C code is a helper function used in...

August 27, 2025 at 06:47 AM

static void find_best_pitch(opus_val32 *xcorr, opus_val16 *y, int len,
                            int max_pitch, int *best_pitch
#ifdef FIXED_POINT
                            , int yshift, opus_val32 maxcorr
#endif
                            )
{
   int i, j;
   opus_val32 Syy=1;
   opus_val16 best_num[2];
   opus_val32 best_den[2];
#ifdef FIXED_POINT
   int xshift;

   xshift = celt_ilog2(maxcorr)-14;
#endif

   best_num[0] = -1;
   best_num[1] = -1;
   best_den[0] = 0;
   best_den[1] = 0;
   best_pitch[0] = 0;
   best_pitch[1] = 1;
   for (j=0;j<len;j++)
      Syy = ADD32(Syy, SHR32(MULT16_16(y[j],y[j]), yshift));
   for (i=0;i<max_pitch;i++)
   {
      if (xcorr[i]>0)
      {
         opus_val16 num;
         opus_val32 xcorr16;
         xcorr16 = EXTRACT16(VSHR32(xcorr[i], xshift));
#ifndef FIXED_POINT
         /* Considering the range of xcorr16, this should avoid both underflows
            and overflows (inf) when squaring xcorr16 */
         xcorr16 *= 1e-12f;
#endif
         num = MULT16_16_Q15(xcorr16,xcorr16);
         if (MULT16_32_Q15(num,best_den[1]) > MULT16_32_Q15(best_num[1],Syy))
         {
            if (MULT16_32_Q15(num,best_den[0]) > MULT16_32_Q15(best_num[0],Syy))
            {
               best_num[1] = best_num[0];
               best_den[1] = best_den[0];
               best_pitch[1] = best_pitch[0];
               best_num[0] = num;
               best_den[0] = Syy;
               best_pitch[0] = i;
            } else {
               best_num[1] = num;
               best_den[1] = Syy;
               best_pitch[1] = i;
            }
         }
      }
      Syy += SHR32(MULT16_16(y[i+len],y[i+len]),yshift) - SHR32(MULT16_16(y[i],y[i]),yshift);
      Syy = MAX32(1, Syy);
   }
}

The provided C code is a helper function used in audio processing, particularly in finding the best pitch candidates to aid in tasks such as pitch detection, audio compression, or speech analysis. Let’s break down what this find_best_pitch function does:

Purpose:

The function identifies the two best pitch candidates (best_pitch[0] and best_pitch[1]) from a cross-correlation analysis of an input signal. It compares all candidates based on specific criteria to determine the strongest periodic matches between segments of the signal.

Parameters:

opus_val32 *xcorr: Precomputed cross-correlation values for the signal, passed as an array.
opus_val16 *y: The signal data for which pitch is being analyzed.
int len: The length of the segment being analyzed.
int max_pitch: The maximum pitch period to consider. It limits the range of pitch candidates.
int *best_pitch: The output array that stores the two best pitch candidates (index values of xcorr).

Optional (FIXED_POINT specific): 6. int yshift: A scaling factor used in fixed-point arithmetic to manage precision. 7. opus_val32 maxcorr: The maximum cross-correlation value, used to normalize computation for fixed-point mode.

Breakdown of Function Logic:

Initialization:
- The function initializes variables like Syy to start with a value of 1.
- Arrays best_num and best_den are initialized to hold numerators and denominators of the scoring metric for the two best pitch candidates.
Compute Signal Energy (Syy):
- Syy is initialized by summing the squared values of the signal y over the segment length len.
- This is part of the normalization process to ensure correlation values are scaled relative to signal energy.
Iterate Over Pitch Candidates:
- The primary loop iterates over all pitch candidates from 0 to max_pitch - 1.
- If the cross-correlation (xcorr[i]) for a pitch candidate is positive, the function proceeds to compute its score.
Score Calculation:
- The function calculates a score for the current pitch candidate, which involves squaring the normalized cross-correlation value (xcorr16) and comparing it to previously identified best scores.
- The score is evaluated using the formula: (num / den) where num is the square of the normalized cross-correlation, and den is the energy Syy.
- If the current score is better than the second-best (best_num[1]), it updates the list of best pitch candidates.
Sliding Window Update:
- The energy Syy is updated in a sliding-window manner as the loop progresses. For the next pitch candidate, it subtracts the squared value of the signal element leaving the window (y[i]) and adds the squared value of the new element entering the window (y[i + len]).
- Syy is clamped to a minimum of 1 to avoid division by zero or overly small values.
Best Pitch Results:
- At the end of the loop, the top two pitch candidates (based on the score) are stored in the best_pitch array.

Key Details:

Fixed-Point and Floating-Point Modes:
- The function is designed to work in both fixed-point and floating-point environments, depending on whether the FIXED_POINT macro is defined.
- In fixed-point mode, additional calculations involving yshift and xshift are used to manage precision and scaling.
Numerical Stability:
- To avoid issues such as overflow, underflow, or division by small numbers, the function clamps or scales values (e.g., Syy = MAX32(1, Syy)).

What Does This Code Achieve?

This code is specifically designed to detect the most likely pitch periods for a given signal based on the precomputed cross-correlation values. It ensures robust detection by normalizing scores and handling edge cases like low energy. Such functionality is commonly found in low-level audio codecs or analysis tools like Opus, where pitch estimation is used for encoding, decoding, or enhancing audio signals.

Generate your own explanations

Download our vscode extension

Read other generated explanations

Built by @thebuilderjr