TL;DR
This paper introduces the "Talk & Run" (TalkR) Speech database, a novel dataset of speech recorded from individuals running at various intensities while talking. They demonstrate that speech analysis can automatically detect exercise intensity levels (specifically, distinguishing between aerobic and anaerobic zones) with promising accuracy, paving the way for unobtrusive, smartphone-based fitness monitoring.
What Problem Does This Paper Address?
Understanding and monitoring exercise intensity is crucial for effective training and preventing overexertion. While heart rate monitors are common, they can be cumbersome. The "Talk Test" offers a subjective, accessible method: if you can talk comfortably, your intensity is likely moderate. However, this subjective assessment is difficult to automate. This paper aims to bridge this gap by investigating whether objective, automatic analysis of speech characteristics during running can reliably detect exercise intensity, specifically differentiating between moderate (aerobic) and high (anaerobic) zones.
Methods
The core of this work is the creation of the TalkR Speech database. Participants ran on a treadmill at different speeds, progressing from a warm-up phase to volitional exhaustion. During this, they were asked to read aloud short texts. Speech was recorded using lavalier microphones. Alongside speech, researchers collected heart rate, perceived exertion (Borg scale), and the "Feeling Scale" (FS) to capture subjective experience.
For the automatic detection task, they employed a speaker-independent classification approach. Speech data was segmented into 4-second chunks. A set of 224 speech features were extracted using the openSMILE toolkit, including Mel-Frequency Cepstral Coefficients (MFCCs), energy, fundamental frequency (F0), and zero-crossing rate, along with their temporal derivatives and statistical functionals. A Support Vector Machine (SVM) with a Radial Basis Function (RBF) kernel was trained to classify speech segments into two classes: low exercise intensity (below the anaerobic threshold) and high exercise intensity (above the anaerobic threshold).
Datasets
The paper introduces the Talk & Run (TalkR) Speech database. This dataset contains speech recordings from 21 healthy young adults (15 female, 6 male) who ran on a treadmill while reading aloud. The recordings were made during different phases of exercise: rest, warm-up, and exercise up to volitional exhaustion. Crucially, the speech content varied during the running phases, allowing for a more robust assessment of automatic detection compared to studies using fixed speech content.
Key Insights and Ideas
- Speech as a proxy for physiological state: The fundamental idea is that physical exertion directly impacts the respiratory system, which in turn affects speech production. This paper explores whether these subtle speech changes can be reliably linked to specific exercise intensity levels.
- Automating the "Talk Test": The research directly addresses the potential to automate the subjective "Talk Test" by using objective speech features. This could democratize fitness monitoring, making it accessible via common devices like smartphones.
- Varying lexical content is key: Unlike previous work that often used fixed phrases, the TalkR database incorporates varied speech content during exercise. This is a significant step towards developing systems that can generalize to real-world conversational scenarios during physical activity.
- Gender differences in speech-based intensity detection: The results highlight a notable performance gap between female and male speakers, suggesting that gender-specific models or feature sets might be necessary for optimal performance.
Key Results
The authors achieved promising results in automatically detecting exercise intensity:
- Female speakers: An unweighted average recall (UAR) of 73.5% was achieved in classifying speech into low vs. high exercise intensity.
- Male speakers: The UAR was 60.0% for male speakers.
These results indicate that it is feasible to distinguish between aerobic and anaerobic exercise intensity levels in speech during running, with better performance for female speakers.
Limitations and Future Work
The authors acknowledge several limitations and outline future research directions:
- Limited male speaker data: The relatively lower performance for male speakers is attributed to the smaller number of male participants (6) compared to female participants (15) in the database.
- Outdoor robustness: The current experiments were conducted in a controlled lab environment. Future work will investigate the robustness of the classifier for outdoor use.
- Expanding classification: The current study focused on a binary classification (low vs. high intensity). Future work aims to investigate more granular classification of exercise intensity levels.
- Feature and algorithm exploration: The authors plan to explore a wider range of speech features and classification algorithms to improve performance.
- Correlation analysis: Investigating correlations between subjective measures (FS, RPE, TT) and objective speech/movement characteristics is a key future direction.
- Phonetic analysis: A more detailed analysis of phonetic features affected by running is planned.
- Database release: The TalkR Speech database is intended to be released for research purposes.
Summary updated