TL;DR
This paper introduces the Multi-Session Audio Research Project (MARP) corpus, a longitudinal dataset designed to investigate factors affecting speaker recognition accuracy. Initial findings reveal that subtle variations in intonation, whispering, and even cross-session testing can significantly degrade performance, highlighting the sensitivity of current systems to real-world acoustic conditions.
What Problem Does This Paper Address?
Speaker Recognition (SR) systems are notoriously susceptible to "mismatch" between training and testing conditions. While factors like noise and channel distortion are well-studied, the impact of more subtle, yet common, variations like changes in a speaker's vocal patterns over time, intonation, or even the act of whispering, are less understood. This paper addresses this gap by introducing the MARP corpus, a rich, longitudinal dataset specifically designed to isolate and quantify the effects of these less-explored factors on speaker identification accuracy. The authors aim to provide a formal citation and share preliminary findings to guide future SR algorithm development and mitigation strategies.
Methods
The core of the SR system used in this study is the Gaussian Mixture Model (GMM) and Universal Background Model (UBM) approach, a standard technique in speaker recognition. The front-end processing involves mel-weighted cepstra and their deltas. What's novel here isn't the SR system itself, but its application to a carefully designed corpus to test specific hypotheses. The experiments involve training SR models on various subsets of the MARP data and testing them under different conditions, meticulously comparing in-session vs. cross-session performance, phonated vs. whispered speech, and text-dependent vs. text-independent scenarios.
Datasets
The paper introduces the Multi-Session Audio Research Project (MARP) Corpus. This is a newly collected, longitudinal dataset comprising 21 sessions over three years with over 60 participants. Each session includes:
- Read spoken sentences (phonetically diverse, with variations in intonation).
- Whispered sentences.
- A read passage.
- Read digits.
- A free-flowing conversation.
The corpus was designed to specifically test the impact of aging, inter-session variability, intonation, whispering, text dependency over time, and read vs. spontaneous speech. For evaluation, the authors also reference the TIMIT corpus for GMM adaptation.
Key Insights and Ideas
- Subtle vocal variations matter: Even in controlled anechoic environments, cross-session testing (across months) leads to a dramatic increase in error rates, suggesting that "micro-register" variations in a speaker's voice over time are a significant challenge.
- Intonation is a powerful cue: Changes in intonation (declarative, exclamatory, interrogative) significantly impact SR accuracy, indicating that formant structure alterations due to pitch changes are important discriminators.
- Whispering is a major hurdle: The performance degradation when dealing with whispered speech is substantial, suggesting that current SR systems heavily rely on characteristics present in normal phonation.
- Text dependency can be a double-edged sword: While text dependency generally improves performance, especially across sessions, the study highlights that SR systems can be "tricked" by phonetic content, sometimes prioritizing a text match from a different speaker over a true speaker match with different text.
Key Results
- Cross-Session Impact: In-session tests averaged 3.11% error, while cross-session tests averaged 20.2% error, a 17% increase.
- Intonation Impact: Intonation mismatch increased error by an average of 9.7% over matched conditions. Exclamatory/interrogative training on declarative tests yielded 10.1% error, while declarative training on exclamatory/interrogative tests resulted in 12.8% error.
- Whispering vs. Phonation: Whispering dramatically impacted performance, with an average error rate of 84.3% when training and testing across whispered and phonated speech.
- Text Dependency: In text-dependent tests, there was 0% error. However, in a "reverse text dependency" scenario (where other speakers' models matched the text but the target speaker's did not), error jumped to 91.9%, indicating that phonetic content can override speaker identity.
Limitations and Future Work
Limitations: * The paper acknowledges that the "reverse text dependency" scenario was designed to specifically test how easily SR systems can be misled by text. * Whispered-on-whispered tests could not be performed initially due to insufficient data, though this was remedied in later sessions.
Future Work: * The MARP corpus is being prepared for distribution via the Linguistic Data Consortium. * Future research will explore the impact of five microphones in different locations, varying degrees of natural noise introduced to headphones (to isolate vocal effects from environmental noise), and the addition of noise to recordings to differentiate noise contamination from voice alterations. * Further investigation into the nature of inter-session variability and its correlates is planned to improve SR reliability over time. * The authors suggest that longer time-frame, non-modal features might be more accurate than current MFCCs for capturing speaker characteristics.
Summary updated