Introduction
High-quality speech transformation, such as pitch shifting and time-scaling, often suffers from artifacts like "phasiness" or a "metallic" quality. Traditional high-end methods like PSOLA or STRAIGHT mitigate these issues by relying on precise fundamental frequency (F0) estimation or pitch marker detection. However, these dependencies make such systems fragile in noisy environments or computationally intensive for real-time applications. This paper introduces the Shape-Invariant Phase vocoder (SHIP), an algorithm designed to achieve high-quality transformations without the need for explicit pitch tracking.
The Problem of Vertical Phase Synchronization
Standard phase vocoders often lose "vertical phase synchronization"—the alignment between harmonics that defines the glottal pulse's shape. When this synchronization is lost, the resulting speech sounds blurry. While pitch-synchronous methods solve this, they are prone to F0 estimation errors. The SHIP algorithm addresses this by preserving the waveform's structure (shape invariance) through a robust, F0-free mechanism.
The SHIP Algorithm
The SHIP algorithm embeds the logic of Synchronized Overlap-Add (SOLA) within a phase vocoder framework. Instead of tracking pitch, it determines the optimal time shift by performing cross-correlation in the spectral domain between the current spectral frame and the previous synthesis frame.
Key technical features include:
- Sinusoidal Masking: A mask is applied to focus correlation on sinusoidal peaks (harmonics), ignoring noise that could skew alignment.
- Voiced/Unvoiced Handling: A two-band model maintains shape invariance below a frequency boundary while applying phase randomization above it to keep unvoiced sounds natural.
- SNR Control: The balance between harmonic and residual components is adjusted during transposition to ensure a natural voice quality.
Experimental Evaluation
The performance of SHIP was evaluated using high-quality recordings from male and female speakers. A subjective listening test involving 19 participants (both experts and non-experts) compared SHIP against established methods like PSOLA, HNM, and STRAIGHT using a 5-level Mean Opinion Score (MOS).
Results and Discussion
Subjective tests indicated that SHIP performed significantly better than PSOLA in most scenarios, particularly for time-scaling and gender-specific transformations. The algorithm proved stable over a wide range of parameters, making it suitable for extreme creative transformations.
Furthermore, SHIP is highly efficient. In real-time tests on a 1.7GHz CPU, it utilized only 10–20% of resources to process 44.1kHz audio. Its latency is minimal, restricted primarily by the analysis window size.
Conclusion
The SHIP algorithm demonstrates that high-quality, shape-invariant speech transformation is possible without the overhead and fragility of pitch tracking. While future work aims to better handle excitation mismatches during extreme transposition, SHIP currently offers a robust and efficient solution for real-time speech processing.
Summary updated