Developer Ditches Phase Vocoder for WSOLA to Fix Robotic Voice in Speech App
A developer building a Voice Design app encountered an unnatural, metallic sound whenever the speech-rate slider was adjusted, tracing the issue to phase blurring caused by librosa's phase vocoder algorithm. The phase vocoder processes audio in the frequency domain and, while mathematically precise, gradually breaks down harmonic phase relationships in human speech, producing a robotic or echoey quality. This degradation is especially noticeable in voice content due to its reliance on formants and harmonic structures, and worsens as the stretch ratio increases. To fix this, the developer implemented WSOLA (Waveform Similarity Overlap-Add) from scratch using NumPy, a time-domain method that avoids phase recalculation entirely by using cross-correlation to find naturally connecting waveform segments. The switch eliminated the robotic artifact, making WSOLA a more suitable choice than the phase vocoder for speech-rate adjustment in voice-focused applications.
This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)
Log in to join the discussion and vote.
Log in