Introduction
Presented at INTERSPEECH 2014, this study from Microsoft Research investigates the transition from traditional Gaussian Mixture Models (GMMs) to Deep Neural Networks (DNNs) in speech recognition. While the performance superiority of DNNs was well-established by this time, the specific nature of these gains—whether they stemmed from inherent robustness to environmental factors or superior discriminative power—remained a subject of inquiry.
Methodology
The researchers compared two primary acoustic models trained on a 400-hour "real-world" dataset comprising Mobile Voice Search (VS) and Short Message Dictation (SMD) data:
- CD-GMM-HMM: A discriminative model utilizing 39-dimension MFCC features, trained with fMPE and bMMI criteria.
- CD-DNN-HMM: A deep network trained on 87-dimension log filter bank (LFB) features with an 11-frame context window.
Both models shared identical training data, decision trees for state tying, and initial alignments to ensure a controlled comparison across Signal-to-Noise Ratio (SNR), device types, and speaking rates.
Phoneme-Level Analysis
A key finding of the study is that DNNs provide a universal improvement in phone discrimination. The Phone Error Rate (PER) was reduced for every single phoneme tested, with relative reductions ranging from 15.6% to 39.8%. Notably, the DNN excelled at "hard" consonants where GMMs traditionally struggled, leading to a more balanced performance profile across the phonetic spectrum.
Environmental Robustness and Channel Mismatch
Despite the overall accuracy gains—a 20.3% relative Word Error Rate Reduction (WERR) for Voice Search and 25.1% for SMD—the study revealed that DNNs are not inherently immune to environmental noise. Performance degradation followed a consistent "1dB rule," where every 1dB drop in SNR resulted in a 1–2% relative performance loss for both models. Similarly, while DNNs outperformed GMMs across all tested mobile devices, the performance gap between the best and worst hardware remained consistent, suggesting that DNNs do not fully resolve channel mismatch issues.
Speaking Rate Sensitivity
The analysis of speaking rate revealed a characteristic "U-shaped" error pattern for both GMM and DNN models. Optimal performance is achieved within a "sweet spot" of approximately 10–12 phones per second. Deviations from this rate—either significantly faster or slower speech—resulted in substantial increases in WER, sometimes as high as 15–30%.
Conclusion
The study concludes that while DNNs represent a significant leap forward in acoustic modeling, they are not a "silver bullet" for all speech recognition challenges. The persistence of sensitivity to noise, channel variation, and speaking rate suggests that future improvements must integrate speech enhancement, channel normalization, and explicit temporal modeling to achieve true environmental robustness.
Summary updated