Bug Report: Redundant Conditional Block for `MCD_mode == "plain"`

[Zijun-Mo]

Description

In the function [average_mcd](https://github.com/modelscope/ClearerVoice-Studio/blob/main/speechscore/scores/mcd.py), the condition

if MCD_mode == "plain":

is checked twice in a row, causing logical redundancy and potential dead code paths.

Here is the problematic section:

if MCD_mode == "plain":
    # pad 0
    if len(loaded_ref_wav) < len(loaded_syn_wav):
        loaded_ref_wav = np.pad(loaded_ref_wav, (0, len(loaded_syn_wav) - len(loaded_ref_wav)))
    else:
        loaded_syn_wav = np.pad(loaded_syn_wav, (0, len(loaded_ref_wav) - len(loaded_syn_wav)))

    # extract MCEP features
    ref_mcep_vec = self.wav2mcep_numpy(loaded_ref_wav, score_rate)
    syn_mcep_vec = self.wav2mcep_numpy(loaded_syn_wav, score_rate)

    if MCD_mode == "plain":  # ← redundant check
        path = []
        for i in range(len(ref_mcep_vec)):
            path.append((i, i))
    elif MCD_mode == "dtw":
        _, path = fastdtw(ref_mcep_vec[:, 1:], syn_mcep_vec[:, 1:], dist=euclidean)
    elif MCD_mode == "dtw_sl":
        cof = len(ref_mcep_vec)/len(syn_mcep_vec) if len(ref_mcep_vec)>len(syn_mcep_vec) else len(syn_mcep_vec)/len(ref_mcep_vec)
        _, path = fastdtw(ref_mcep_vec[:, 1:], syn_mcep_vec[:, 1:], dist=euclidean)

Expected Behavior

The outer if MCD_mode == "plain": should likely be removed or changed to allow all three modes (plain, dtw, dtw_sl) to follow a similar processing pipeline (padding + feature extraction).
Currently, the inner conditions for "dtw" and "dtw_sl" will never be executed, because they are nested inside a block that only runs if MCD_mode == "plain".

Actual Behavior

• When MCD_mode is "dtw" or "dtw_sl", the code inside the inner elif blocks is never reached.
• As a result, DTW-based MCD computation does not execute as intended.

Proposed Fix

Move the MCEP extraction part outside of the "plain" condition and structure it like this:

# pad shorter wav
if len(loaded_ref_wav) < len(loaded_syn_wav):
    loaded_ref_wav = np.pad(loaded_ref_wav, (0, len(loaded_syn_wav) - len(loaded_ref_wav)))
else:
    loaded_syn_wav = np.pad(loaded_syn_wav, (0, len(loaded_ref_wav) - len(loaded_syn_wav)))

# extract features
ref_mcep_vec = self.wav2mcep_numpy(loaded_ref_wav, score_rate)
syn_mcep_vec = self.wav2mcep_numpy(loaded_syn_wav, score_rate)

# choose path strategy
if MCD_mode == "plain":
    path = [(i, i) for i in range(len(ref_mcep_vec))]
elif MCD_mode == "dtw":
    _, path = fastdtw(ref_mcep_vec[:, 1:], syn_mcep_vec[:, 1:], dist=euclidean)
elif MCD_mode == "dtw_sl":
    cof = len(ref_mcep_vec) / len(syn_mcep_vec) if len(ref_mcep_vec) > len(syn_mcep_vec) else len(syn_mcep_vec) / len(ref_mcep_vec)
    _, path = fastdtw(ref_mcep_vec[:, 1:], syn_mcep_vec[:, 1:], dist=euclidean)

Environment

• Repository: [modelscope/ClearerVoice-Studio](https://github.com/modelscope/ClearerVoice-Studio)
• File: speechscore/scores/mcd.py
• Function: average_mcd
• Python version: 3.11.13
• OS: Ubuntu 22.04

Additional Context

This bug likely causes incorrect MCD computation for "dtw" and "dtw_sl" modes, leading to invalid or nan results during evaluation.