In 2005, a team from the University of Waterloo presented a system at Interspeech that used a piezoelectric contact microphone placed on the throat to capture speech in a factory floor setting. Their hidden Markov model (HMM) recognizer achieved a word error rate of 12% under 85 dB of background noise—a figure that air microphones could not match below 40%. This result was not an outlier; throughout the mid-2000s, contact microphones quietly enabled speech recognition in environments where traditional acoustic sensors failed.
What Are Contact Microphones?
Contact microphones transduce mechanical vibrations directly from a solid surface—skin, bone, or a structural element—rather than from air pressure waves. The most common types in the mid-2000s were piezoelectric disks and accelerometer-based sensors. When placed on the neck, cheek, or behind the ear, they capture the speaker’s voice through bone conduction and tissue vibration, largely rejecting airborne noise. This gave them a unique advantage in high-noise scenarios: the signal-to-noise ratio (SNR) of a contact microphone could remain above 20 dB even when the ambient noise level exceeded 100 dB SPL.
Researchers at the time documented several acoustic properties of contact speech. The frequency range was typically limited to 300 Hz–4 kHz, compared to 100 Hz–8 kHz for a good air microphone, but the spectral envelope retained enough phonetic information for recognition. The amplitude response also varied with placement; a sensor on the thyroid cartilage captured different harmonics than one on the maxilla.
Integration with Mid-2000s Speech Recognition Systems
The dominant paradigm for speech recognition in the mid-2000s was the HMM with Gaussian mixture models (GMMs), fed by Mel-frequency cepstral coefficients (MFCCs) or perceptual linear prediction (PLP) features. Contact microphone signals required adaptation because their spectral tilt and formant structure differed from air-conducted speech. Some groups developed specialized front-end processing that emphasized lower-frequency bands or used linear predictive coding (LPC) residuals.
The acoustic features extracted from contact microphones often relied on time-domain measures such as zero-crossing rate, a method that has been largely forgotten but still holds value for certain applications—as discussed in a previous post on Zero Crossing Analysis: The Forgotten Acoustic Feature of Mid-2000s Speech Recognition. In particular, zero-crossing density helped distinguish voiced from unvoiced segments in contact speech, where the usual energy-based voicing detectors performed poorly due to the lack of strong fundamental frequency harmonics.
A typical mid-2000s system pipeline for contact microphone recognition looked like this:
- Signal acquisition via a piezoelectric sensor, preamplified and bandpass filtered (100 Hz–5 kHz).
- Segmentation using a combination of zero-crossing rate and short-time energy, with thresholds tuned to the contact sensor’s dynamic range.
- Feature extraction: 13 MFCCs (with cepstral mean subtraction) plus delta and acceleration coefficients, computed over 25 ms frames with 10 ms shift.
- Recognition using a tied-state triphone HMM trained on a corpus of contact speech recordings, often augmented by artificial noise injection.
One notable challenge was the lack of large, publicly available contact speech databases. Most groups recorded their own small corpora—typically 5–10 speakers reading phonetically balanced sentences in quiet and noisy conditions. This limited the generalizability of the models, but the relative simplicity of the contact channel meant that speaker adaptation techniques like MLLR (maximum likelihood linear regression) worked well even with sparse data.
Applications Beyond Noise Reduction
Contact microphones found use in several niche areas that aligned with the blog’s focus on low-resource languages and biometrics.
Field Recordings for Endangered Languages
Linguists working with endangered languages in remote, noisy environments (e.g., marketplaces, windy villages) adopted contact microphones to capture clean speech samples. In 2006, a documentation project for the Ainu language in Japan used a throat-contact microphone to record elderly speakers while a diesel generator ran nearby. The resulting recordings were later used to train a small-vocabulary recognizer for language preservation.
Speaker Identification and Biometrics
Because contact microphones capture the speaker’s unique bone-conduction signature, they offered a secondary channel for speaker verification. A 2007 paper from IDIAP Research Institute combined a contact sensor with a conventional air microphone in a score-level fusion framework, reducing equal error rates from 5.2% to 2.8% on a 50-speaker task. The contact channel was particularly robust against replay attacks and background speech.
The same principle was applied to medical voice monitoring. Patients with vocal fold disorders wore contact microphones during daily activities to log voice use patterns, and researchers used HMM-based classifiers to detect vocal fatigue from the contact signal alone.
Challenges and Legacy
Despite their advantages, contact microphones never became mainstream in consumer speech recognition. The need for physical contact was a practical barrier—users did not want to wear a sensor on their neck. Calibration was also finicky: a slight shift in placement changed the frequency response, requiring either robust feature normalization or frequent retraining. Moreover, the rise of deep learning in the late 2000s and early 2010s allowed air microphones to achieve comparable noise robustness through massive training data and multi-channel processing, reducing the incentive for specialized hardware.
Nevertheless, the mid-2000s research left a lasting impact on the field. The techniques for front-end adaptation, noise-robust feature extraction, and cross-channel fusion developed for contact microphones were later repurposed for other non-acoustic sensors, such as bone-conduction headphones and in-ear microphones. The databases of contact speech, though small, continue to be used as benchmarks for extreme noise conditions.
Today, contact microphones survive in specialized domains: aviation headsets where pilots communicate in roaring cockpits, law enforcement body-worn sensors for covert recording, and medical voice dosimetry. The mid-2000s work demonstrated that when the acoustic environment is hostile, the most reliable signal is often the one that bypasses the air entirely.