Abstract
Mandarin is a tonal language where pitch is a critical acoustic cue for distinguishing monosyllabic digits. However, traditional Hidden Markov Model (HMM) based speech recognition systems struggle with "unvoiced" segments that lack a clear pitch, often relying on random placeholder values that cause statistical confusion. This paper introduces a distribution normalization technique to prevent unvoiced noise from confusing the model and adds a "Voice Level" feature to better distinguish speech patterns. Our approach results in up to a 15% reduction in sentence error rates (SER) for Mandarin Digit String Recognition (MDSR) tasks.
Introduction
Mandarin Digit String Recognition (MDSR) is notoriously difficult because Mandarin is a tonal language composed of monosyllables. Many digits sound very similar (e.g., ‘2’ and ‘8’, or ‘0’ and ‘6’), making pitch a vital cue for distinguishing them.
A significant technical hurdle exists: HMMs require a continuous stream of data. While "voiced" (V) sounds like vowels have a clear pitch, "unvoiced" (UV) sounds like 's' or 'f' do not. Traditionally, random values are used for UV frames, leading to two major issues:
- Statistical Confusion: Random values often overlap with the range of real pitch values, tricking the model into misidentifying unvoiced segments as voiced ones.
- Likelihood Spikes: Wildly unexpected random values can cause "abnormal likelihoods," leading the decoder to prematurely prune correct paths.
Proposed Methodology
1. Distribution Normalization
To address statistical confusion, the authors propose a linear transformation for UV frames: $$F'(UV) = a \cdot F(UV) + b$$ This shifts the distribution of placeholder values to have the same variance as real pitch features but separated by a specific "distance factor" ($d$). This ensures UV values are mapped to a range that minimizes overlap with voiced pitch values, effectively embedding V/UV decision information into the feature stream.
2. Voice Level (VL) Features
A new 2-dimensional feature called Voice Level is introduced. VL measures the "intensity of periodicity"—how "voice-like" a segment is—calculated as the difference between the maximum and minimum of the summed cross-correlation during pitch estimation. By adding static VL and its first derivative (delta) to the feature vector, the system better distinguishes speech from noise.
Experimental Evaluation
The system was trained on 156,975 clean Mandarin utterances and evaluated on 2,000 utterances (4-8 digits long) under various noise conditions, including car noise at 5dB to 20dB SNR.
Optimization of the Distance Factor
The distance factor ($d$) between voiced and unvoiced distributions is critical. If $d=0$, distributions overlap too much; if $|d|$ is too large, the system becomes a "hard decision" model with no tolerance for V/UV errors. Empirical testing found an optimum at $d = -1.2$, achieving an average SER of 11.03%.
Key Results and Insights
The proposed methods showed cumulative improvements over a strong baseline using Minimum Phone Error (MPE) training:
- Baseline (Pitch + MPE): 12.49% Average SER.
- With Distribution Normalization ($d=-1.2$): 11.03% SER.
- With Normalization + VL Features: 10.83% SER.
This represents a 13.3% Error Rate Reduction (ERR) on free-length strings and a 15.1% ERR on a specific 6-digit task. By normalizing the UV distribution, the authors "pump" front-end V/UV decisions into the continuous pitch feature, allowing the HMM to "soften" these decisions during decoding.
Conclusion
This work demonstrates that careful handling of unvoiced pitch features and the inclusion of periodicity strength (VL) significantly improves Mandarin digit recognition. While effective for small vocabularies, future work should explore scaling these techniques to Large Vocabulary Continuous Speech Recognition (LVCSR) and automating the tuning of the distance factor $d$.
Summary updated