Abstract:Accurate pitch extraction from speech is important but challenging problem for speech synthesis. However, the additive nature and long-term suprasegmental property of pitch features have not been fully exploited in most of the existing pitch estimators as they are operated frame by frame. As a result, they would cause some inherent discontinuities, such as double/half F0 errors and unvoiced/voiced (U/V) error. This would adversely affect the quality of synthetic speech as well as the expressiveness of the prosody information. In this paper, we explore the novel use of multi-tasks (Task 1: U/V; Task 2: Pitch) bidirectional long short-term memory recurrent neural network (BLSTM) to model the pitch and voicing decision simultaneously in a unified framework. The features used in this study are extracted from the frequency domain. We compute the log-frequency power spectrogram and then normalize to the long-term speech spectrum to attenuate noises. A filter is then used to enhance the harmonicity. Experiments show that the proposed approach substantially outperforms RAPT, which behaves the best in clean condition. Besides, our proposed approach can even work well with a certain level of background noise.

Research on real-time detection technology of Chinese voiced speech pitch

An Algorithm for Realtime High Resolution Octave Analysis Based on Multirate Signal Processing Theory

Optimization of Pitch Preprocessing in TETRA Speech Encoder

A Power Spectrum Reprocessing Algorithm for Pitch Detection of Speech

Improved Algorithm for Pitch Detection and Harmonic Separation

The Research on Pitch Extraction Method for Voice Activity Detection Based on Periodic Decomposition

Pitch Detection Method Based on Hilbert-Huang Transform for Speech Signals

Improved Algorithm for Pitch Period Detection

Detection of Time Varying Pitch in Tonal Languages: an Approach Based on Ensemble Empirical Mode Decomposition

Pitch Extraction Based on Wavelet Transformation and Linear Prediction

Speech endpoint detection based on frequency domain and time domain analyses

Detectionoftimevaryingpitchintonallanguages: Anapproachbasedonensembleempirical Modedecomposition ∗

Voice Activity Detection Algorithm Based on Long-Term Pitch Information

A New Robust Pitch Determination Algorithm for Telephone Speech

Multi-Pitch Detection for Co-Channel Speech Utilizing Frequency Channel Piecewise Integration and Morphological Feedback Verification Tracking

A Pitch Period Detection Algorithm Using Time and Frequency Analyses

Real-time Pitch Tracking Based on C

Latent time-frequency component analysis: A novel pitch-based approach for singing voice separation

A New Method for Accurate Pitch Period Measurement

High performance mandarin pitch estimation

An initial research: Towards accurate pitch extraction for speech synthesis based on BLSTM