yskim3271/Throat_and_Acoustic_Pairing_Speech_Dataset

TAPS: Throat and Acoustic Paired Speech Dataset

1. DATASET SUMMARY

The Throat and Acoustic Paired Speech (TAPS) dataset is a standardized corpus designed for deep learning-based speech enhancement, specifically targeting throat microphone recordings. Throat microphones effectively suppress background noise but suffer from high-frequency attenuation due to the low-pass filtering effect of the skin and tissue. The dataset provides paired recordings from 60 native Korean speakers, captured simultaneously using a throat microphone (accelerometer-based) and an acoustic microphone. This dataset facilitates speech enhancement research by enabling the development of models that recover lost high-frequency components and improve intelligibility. Additionally, we introduce a mismatch correction technique to align signals from the two microphones, which enhances model training.

---

2. Dataset Usage

To use the TAPS dataset, follow the steps below:

2.1 Loading the dataset

You can load the dataset from Hugging Face as follows:

from datasets import load_dataset
dataset = load_dataset("yskim3271/Throat_and_Acoustic_Pairing_Speech_Dataset")
print(dataset)

Example output:

 DatasetDict({
     train: Dataset({
         features: ['gender', 'speaker_id', 'sentence_id', 'text', 'duration', 'audio.throat_microphone', 'audio.acoustic_microphone'],
         num_rows: 4000
     })
     dev: Dataset({
         features: ['gender', 'speaker_id', 'sentence_id', 'text', 'duration', 'audio.throat_microphone', 'audio.acoustic_microphone'],
         num_rows: 1000
     })
     test: Dataset({
         features: ['gender', 'speaker_id', 'sentence_id', 'text', 'duration', 'audio.throat_microphone', 'audio.acoustic_microphone'],
         num_rows: 1000
     })
 })

2.2 Accessing a sample

Each dataset entry consists of metadata and paired audio recordings. You can access a sample as follows:

sample = dataset["train"][0]  # Get the first sample
print(f"Gender: {sample['gender']}")
print(f"Speaker ID: {sample['speaker_id']}")
print(f"Sentence ID: {sample['sentence_id']}")
print(f"Text: {sample['text']}")
print(f"Duration: {sample['duration']} sec")
print(f"Throat Microphone Audio Path: {sample['audio.throat_microphone']['path']}")
print(f"Acoustic Microphone Audio Path: {sample['audio.acoustic_microphone']['path']}")

2.3 Using the with_normalized_text Configuration

The dataset provides an additional configuration that includes normalized text transcriptions for the test set.

Loading the configuration:

from datasets import load_dataset

# Load with normalized_text config
dataset = load_dataset(
    "yskim3271/Throat_and_Acoustic_Pairing_Speech_Dataset",
    name="with_normalized_text"
)

Example output:

DatasetDict({
    train: Dataset({
        features: ['gender', 'speaker_id', 'sentence_id', 'text', 'duration',
                   'audio.throat_microphone', 'audio.acoustic_microphone', 'normalized_text'],
        num_rows: 4000
    })
    dev: Dataset({
        features: ['gender', 'speaker_id', 'sentence_id', 'text', 'duration',
                   'audio.throat_microphone', 'audio.acoustic_microphone', 'normalized_text'],
        num_rows: 1000
    })
    test: Dataset({
        features: ['gender', 'speaker_id', 'sentence_id', 'text', 'duration',
                   'audio.throat_microphone', 'audio.acoustic_microphone', 'normalized_text'],
        num_rows: 1000
    })
})

Normalized Text Details:

• Test set: Contains normalized Korean text with numbers spelled out (e.g., "7일" → "칠 일", "120명" → "백 이십 명") and punctuation normalized (quotation marks removed)
• Train/Dev sets: Contains empty strings ("") for schema consistency

Example usage:

sample = dataset["test"][0]
print(f"Original text: {sample['text']}")
print(f"Normalized text: {sample['normalized_text']}")

# Output:
# Original text: 부산시는 7일 오후 시민들이 안전하게 대중교통을 이용하기 위해 '대중교통 혼잡도 안전관리 대책회의'를 열고 정책 점검에 나섰다
# Normalized text: 부산시는 칠 일 오후 시민들이 안전하게 대중교통을 이용하기 위해 대중교통 혼잡도 안전관리 대책회의를 열고 정책 점검에 나섰다

---

3. Links and Details

• Project website: Link
• Point of contact: Yunsik Kim ([email protected])
• Collected by: Intelligent Semiconductor and Wearable Devices (ISWD) of the Pohang University of Science and Technology (POSTECH)
• Language: Korean
• Download size: 7.03 GB
• Total audio duration: 15.3 hours
• Number of speech utterances: 6,000

---

4. Citataion

The BibTeX entry for the dataset is currently being prepared.

---

5. DATASET STRUCTURE & STATISTICS

• Training Set (40 speakers, 4,000 utterances, 10.2 hours)
• Development Set (10 speakers, 1,000 utterances, 2.5 hours)
• Test Set (10 speakers, 1,000 utterances, 2.6 hours)
• Each set is gender-balanced (50% male, 50% female).
• No speaker overlap across train/dev/test sets.

Dataset Type	Train	Dev	Test
Number of Speakers	40	10	10
Number of male speakers	20	5	5
Mean / standard deviation of the speaker age	28.5 / 7.3	25.6 / 3.0	26.2 / 1.4
Number of utterances	4,000	1,000	1,000
Total length of utterances (hours)	10.2	2.5	2.6
Max / average / min length of utterances (s)	26.3 / 9.1 / 3.2	17.9 / 9.0 / 3.3	16.6 / 9.3 / 4.2

---

6. DATA FIELDS

6.1 Default Configuration

Each dataset entry contains:

• gender: Speaker's gender (male/female).
• speaker_id: Unique speaker identifier (e.g., "p01").
• sentence_id: Utterance index (e.g., "u30").
• text: Transcription (provided only for test set).
• duration: Length of the audio sample.
• audio.throat_microphone: Throat microphone signal.
• audio.acoustic_microphone: Acoustic microphone signal.

6.2 with_normalized_text Configuration

In addition to all fields from the default configuration, this configuration includes:

• normalized_text: Normalized transcription with:
  • Numbers spelled out in Korean (e.g., "7" → "칠", "120" → "백 이십")
  • Normalized punctuation (quotation marks removed)
  • Available only for test set (1,000 utterances)

---

7. DATASET CREATION

7.1 Hardware System for Audio Data Collection

The hardware system simultaneously records signals from a throat microphone and an acoustic microphone, ensuring synchronization.

• Throat microphone: The TDK IIM-42652 MEMS accelerometer captures neck surface vibrations (8 kHz, 16-bit resolution).

  

• Acoustic microphone: The CUI Devices CMM-4030D-261 MEMS microphone records audio (16 kHz, 24-bit resolution) and is integrated into a peripheral board.

  

• MCU and data transmission: The STM32F301C8T6 MCU processes signals via SPI (throat microphone) and I²S (acoustic microphone). Data is transmitted to a laptop in real-time through UART communication.

  

7.2 Sensors Positioning and Recording Environment

• Throat microphone placement: Attached to the supraglottic area of the neck.
• Acoustic microphone position: 30 cm in front of the speaker.
• Recording conditions: Conducted in a controlled, semi-soundproof environment to minimize ambient noise.
• Head rest: A headrest was used to maintain a consistent head position during recording.
• Nylon filter: A nylon pop filter was placed between the speaker and the acoustic microphone to minimize plosive sounds.
• Scripts for Utterances: Sentences were displayed on a screen for participants to read.

  

7.3 Python-based Software for Data Recordings

The custom-built software facilitates real-time data recording, monitoring, and synchronization of throat and acoustic microphone signals.

• Interface overview: The software displays live waveforms, spectrograms, and synchronization metrics (e.g., SNR, shift values).
• Shift analysis: Visualizations include a shift plot to monitor synchronization between the microphones and a shift histogram for statistical analysis.
• Recording control: Users can manage recordings using controls for file navigation (e.g., Prev, Next, Skip).
• Real-time feedback: Signal quality is monitored with metrics like SNR and synchronization shifts.

7.4 Recorded Audio Data Post-Processing

• Noise reduction: Applied Demucs model to suppress background noise in the acoustic microphone recordings.
• Mismatch correction: Aligned throat and acoustic microphone signals using cross-correlation.
• Silent segment trimming: Removed leading/trailing silence.

7.5 Personal and Sensitive Information

• No personally identifiable information is included.
• Ethical approval: Institutional Review Board (IRB) approval from POSTECH.
• Consent: All participants provided written informed consent.

---

8. POTENTIAL APPLICATIONS OF THE DATASET

The TAPS dataset enables various speech processing tasks, including:

• Speech enhancement: Improving the intelligibility and quality of throat microphone recordings by recovering high-frequency components.
• Automatic speech recognition (ASR): Enhancing throat microphone speech for better transcription accuracy in noisy environments.
• Speaker Verification: Exploring the effectiveness of throat microphone recordings for identity verification in challenging acoustic environments.