A comparative study on AM and FM features

Yotaro Kubo, Shigeki Okawa, Akira Kurematsu, Katsuhiko ShiraiInterspeech 2008

TL;DR

This paper investigates the complementary nature of Amplitude Modulation (AM) and Frequency Modulation (FM) features for robust speech recognition. By comparing AM and FM classifiers in noisy conditions and analyzing their statistical independence, the authors demonstrate that combining these features leads to improved performance, particularly in challenging acoustic environments.

What Problem Does This Paper Address?

Speech recognition systems often struggle with noisy environments, where the acoustic signal is corrupted. While traditional speech features capture aspects of the signal's energy (related to AM), they may not fully exploit the information present in how the signal's frequency changes over time (FM). Previous work has suggested that combining AM and FM features could enhance robustness, but a thorough evaluation of their complementarity and independence, especially under various noise conditions, was lacking. This paper aims to fill that gap by systematically comparing AM and FM features and their combined effectiveness.

Methods

The authors employ a speech recognition system that utilizes two distinct classifiers: one for AM features and one for FM features.

  • AM Classifier (HATS): This method, introduced by Chen et al., uses a multi-layer perceptron (MLP) trained with a "matched filter" approach. Input speech is first passed through a Bark filterbank. Then, for each channel, an MLP (MLP-OL) is used to extract "AM emphasis" by calculating the energy of the filterbank outputs. These outputs are then fed into a final MLP for monophone classification.
  • FM Classifier: Similar to the AM classifier, the input speech is processed through a Bark filterbank. For FM extraction, the authors define the Logarithmic Pseudo-Instantaneous Frequency (LPIF) based on the time interval between zero-crossing points of the signal within each filterbank channel. This LPIF is then processed by an MLP-OL to extract "FM emphasis," followed by a monophone classification MLP.

The core idea is that the MLP-OL acts as a "matched filter" that can learn to emphasize specific modulation components. The authors then evaluate the performance of these classifiers individually and in combination. To assess their complementarity, they conduct speech recognition experiments under various artificial noise conditions. To further support their claims, they also perform a statistical analysis to measure the independency between the AM and FM features.

Datasets

  • Training Data: The authors used the CENSREC-1 dataset, which is a Japanese translation of the AURORA-2 dataset. This dataset comprises 8,440 utterances from 110 speakers for clean speech training.
  • Testing Data: The test set also utilized CENSREC-1, consisting of 1001 utterances from 104 speakers. Artificial noises were added to this clean speech data at a 10 dB Signal-to-Noise Ratio (SNR).

Key Insights and Ideas

  • Complementarity in Noise: The AM and FM classifiers exhibit significantly different error rates depending on the type of noise. This suggests that they are sensitive to different aspects of the corrupted signal and thus offer complementary information.
  • Statistical Independence: The statistical analysis reveals that AM and FM features are largely independent of each other, especially when signals are corrupted by realistic environmental noise. This independence is crucial because it implies that combining them is unlikely to result in redundant information.
  • FM's Advantage in Full-Range Noise: The FM analysis shows particular advantages in dealing with full-range noises, while AM analysis can be degraded by such conditions.
  • AM's Weakness in Burst Noise: Conversely, the AM analysis is often degraded by full-range burst noise, where FM features tend to perform better.
  • "Matched Filter" for Modulation: The use of MLP-OLs as "matched filters" is a key technique for effectively extracting and emphasizing specific AM or FM modulation components from the filterbank outputs.

Key Results

  • The paper reports that their combined AM/FM approach (AFMC) achieved a significant reduction in word error rate (43.6%) at an SNR of 10 dB in noisy conditions, outperforming previous AM-FM methods.
  • Table 1 shows specific word accuracies for AM and FM classifiers under different noise types at 10 dB SNR:
  • White Noise (wn): AM: 45.53%, FM: 49.46%
    • Band-pass filtered white noise (bpf_wn): AM: 35.31%, FM: 2.89% (FM performs poorly here)
    • Burst noise (burst_wn): AM: 49.12%, FM: 62.90% (FM performs better)
    • Band-pass filtered burst noise (burst_bpf_wn): AM: 62.90%, FM: 34.66% (AM performs better)
    • These results clearly illustrate the complementary nature of the features across different noise types.

Limitations and Future Work

  • Subband Dependency: While the authors claim statistical independence, their distance matrices (Figure 4) suggest that AM features are highly dependent on each other within subbands. They also note that AM and FM features are "slightly dependent on the FM features in the neighbouring subband." This indicates that perfect independence might not be achieved, especially at the subband level.
  • Artificial Noise: The experiments were conducted using artificially generated noise. While this allows for controlled evaluation, it may not perfectly replicate the complexity of real-world noise environments.
  • Future Work: The paper implicitly suggests that further exploration into how to optimally combine AM and FM features, potentially by weighting them based on noise characteristics or using more sophisticated fusion techniques, could lead to even greater improvements. The observed subband dependencies might also be a target for future investigation.

Summary updated

BibTeX
@inproceedings{kubo2008comparative,
  author = {Yotaro Kubo and Shigeki Okawa and Akira Kurematsu and Katsuhiko Shirai},
  title = {{A comparative study on AM and FM features}},
  booktitle = {Interspeech},
  year = {2008},
}
Figure