Adaptive-order fractional Fourier transform features for speech recognition

Hui Yin, Xiang Xie, Jingming KuangInterspeech 2008

TL;DR

This paper introduces a novel speech recognition feature by combining the traditional Mel-frequency cepstral coefficients (MFCC) with the adaptive fractional Fourier transform (FrFT). By adaptively determining the optimal FrFT order for each speech frame, they achieve improved recognition accuracy, particularly in noisy environments.

What Problem Does This Paper Address?

Traditional speech recognition systems heavily rely on Mel-frequency cepstral coefficients (MFCCs), which model speech based on its frequency characteristics. However, speech is inherently non-stationary, meaning its frequency content changes over time due to factors like intonation. While MFCCs capture some of this, they are based on a linear frequency scale and a fixed transform (the Discrete Fourier Transform or DFT). The authors propose that a more flexible time-frequency analysis, like the fractional Fourier transform (FrFT), could better capture the dynamic nature of speech, especially its frequency modulation characteristics. The challenge with FrFT is that its performance is highly dependent on a parameter called the "order," and finding the optimal order for each speech segment has been computationally expensive and lacked a robust method.

Methods

The core innovation is the FrFT-MFCC feature extraction method. Instead of using the standard DFT in the initial step of MFCC computation, they employ the Discrete FrFT (DFrFT). The crucial part is how they determine the optimal order ($\rho$) for the FrFT for each frame. They use the ambiguity function of the signal to adaptively find this order. The ambiguity function provides a measure of how well a signal matches a specific time-frequency representation, and by analyzing its properties, they can infer the optimal FrFT order. Once the FrFT is computed with the adaptive order, the rest of the MFCC pipeline (Mel filterbanks, log energies, and Discrete Cosine Transform or DCT) is applied to these FrFT-based features.

Datasets

  • Chinese Mandarin Digits Database: This database, collected at Beijing Institute of Technology, was used for experiments. It includes:
  • Isolated digits.
  • Short utterances (3-5 connected digits).
  • The corpus was recorded by 125 speakers (63 male, 62 female) multiple times.
  • Speech was sampled at 16kHz with 16-bit quantization.
  • Noisy Speech Data: For noisy experiments, three types of noises were added to the clean isolated digits: white noise, Volvo noise, and babble noise, across five Signal-to-Noise Ratio (SNR) levels (0dB, 5dB, 10dB, 15dB, 20dB).

Key Insights and Ideas

  • Speech as Chirp-like Signals: The paper posits that speech signals, especially within a pitch period, exhibit characteristics that can be better modeled by frequency-modulated (chirp-like) signals, which FrFT is well-suited to analyze.
  • Adaptive Order is Key: The performance of FrFT is highly sensitive to its order. The paper demonstrates that adaptively finding this order for each frame, rather than using a fixed order, significantly improves recognition.
  • Ambiguity Function for Order Selection: The ambiguity function offers a computationally more efficient way to determine the optimal FrFT order compared to some existing methods, making adaptive order selection more feasible.
  • FrFT-MFCC Outperforms MFCC in Specific Noises: The proposed FrFT-MFCC features show a notable advantage over traditional MFCCs, particularly in babble noise conditions, suggesting its potential for environments with multiple speakers.

Key Results

  • Clean Digits: FrFT-MFCC achieved a relative improvement in Word Error Rate (WER) of 15.4% for isolated digits (from 1.43% to 1.21%) and 7.0% for connected digits (from 5.61% to 5.22%) compared to MFCC.
  • Noisy Digits:
    • In white and babble noise conditions at high SNR levels, FrFT-MFCC outperformed traditional MFCC.
  • A significant relative WER drop of 41% was observed in babble noise at 20dB (from 15.2% to 8.97%).
  • However, MFCC performed better in Volvo noise, which is a low-pass noise, suggesting FrFT might not be universally beneficial.
  • Specific Digit Improvements: Confusion matrices revealed that FrFT-MFCC significantly reduced WERs for digits "qil", "jiu3", and "er4", indicating better discrimination for these specific sounds.

Limitations and Future Work

  • Computational Complexity: While the ambiguity function method for order selection is presented as more efficient than some alternatives, the authors acknowledge that it is still computationally more intensive than traditional FFT. They highlight the need for faster and more accurate order determination algorithms.
  • Noise Dependency: The performance of FrFT-MFCC varies with noise type. It excels in babble noise but is less effective in Volvo noise, suggesting further research into noise-robust adaptive order selection or feature combinations.
  • Future Work: The authors suggest exploring FrFT for other speech processing tasks like speech enhancement and speaker recognition. They also propose combining FrFT with human auditory models (e.g., bark bands) for potentially even better performance.

Summary updated

BibTeX
@inproceedings{yin2008adaptive,
  author = {Hui Yin and Xiang Xie and Jingming Kuang},
  title = {{Adaptive-order fractional Fourier transform features for speech recognition}},
  booktitle = {Interspeech},
  year = {2008},
}
Figure