Abstract
This study investigates the comparative effectiveness of Multi-Task Learning (MTL) and Adversarial Learning (AL) in improving Automatic Speech Recognition (ASR) by managing speaker variability. While MTL encourages the model to identify speakers to aid feature extraction, AL forces the model to ignore speaker identity to achieve speaker-invariant representations. Through experiments on the Wall Street Journal dataset using a Gated Convolutional Neural Network, we find that deep models naturally develop speaker invariance. Furthermore, while both paradigms offer similar regularization benefits, AL excels in semi-supervised settings by leveraging untranscribed data with speaker labels to improve acoustic modeling.
Introduction
In Automatic Speech Recognition (ASR), speaker variability is a major hurdle. A robust acoustic model should ideally focus only on linguistic content, remaining invariant to the identity of the speaker. Historically, two diametrically opposed strategies have been used to handle this:
- Multi-Task Learning (MTL): Training the model to recognize both text and speaker identity, under the assumption that speaker-related features aid in understanding speech patterns.
- Adversarial Learning (AL): Training the model to recognize text while actively preventing a sub-network from identifying the speaker, forcing the learning of speaker-invariant representations.
Because these methods are mathematically incompatible—one minimizes speaker classification error while the other maximizes it via gradient reversal—this paper provides a controlled empirical comparison to determine which approach better serves ASR performance.
Methods
The researchers utilized a 17-layer Gated Convolutional Neural Network (Gated ConvNet) architecture. The model features a shared encoder ($E_r$) that bifurcates into two branches: a Transcription Branch ($D_y$) predicting letters via the AutoSeg Criterion (ASG), and a Speaker Branch ($D_s$) predicting speaker identity.
The speaker branch was "forked" at three different depths: IN (after 2 layers), MID (after 8 layers), and OUT (after 15 layers). In the MTL configuration, the sum of both losses was minimized. In the AL configuration, a minimax game was employed where the encoder attempted to "fool" the speaker classifier by maximizing its loss through gradient reversal.
Experiments were conducted primarily on the Wall Street Journal (WSJ) dataset (81.5 hours). For semi-supervised AL, an additional 35 hours of LibriSpeech data was used, providing speaker labels without text transcriptions.
Results and Key Insights
Natural Invariance in Deep Models
A striking finding is that standard ASR models naturally become speaker-invariant as they get deeper. In the baseline model, speaker classification accuracy plummeted from approximately 80% at early layers to 20% at the final layers, suggesting the network inherently filters out speaker identity as "noise" to improve transcription accuracy.
Performance Comparison
Surprisingly, both MTL and AL yielded similar performance gains on the standard WSJ dataset. Both methods appear to act as different forms of regularization, preventing overfitting by constraining model parameters. The baseline achieved a Word Error Rate (WER) of 5.9% on the Nov92 evaluation set, while AL at the "OUT" layer reached 5.6%.
The Semi-Supervised Advantage
The primary advantage of Adversarial Learning emerged in the semi-supervised setting. By utilizing audio with known speakers but no transcripts, the model was pushed toward greater invariance, resulting in a consistent improvement in Letter Error Rate (LER), dropping from the baseline's 4.9% to 4.6%.
Conclusion
While the impact of these techniques on high-resource datasets like WSJ is relatively minor—likely because deep models already develop natural invariance—the results for semi-supervised learning are promising. Future work should focus on applying these adversarial techniques to low-resource languages, where transcribed data is scarce but speaker-labeled audio is more readily available. Additionally, joint optimization of acoustic and language models may help translate LER improvements more effectively into final word accuracy.
Summary updated