Datasets:
Tasks:
Automatic Speech Recognition
Formats:
parquet
Sub-tasks:
audio-intent-classification
Languages:
English
Size:
1K - 10K
License:
| language: | |
| - en | |
| license: mit | |
| task_categories: | |
| - automatic-speech-recognition | |
| task_ids: | |
| - audio-intent-classification | |
| pretty_name: Throat Microphone Dataset | |
| size_categories: | |
| - 1K<n<10K | |
| tags: | |
| - audio | |
| - speech | |
| - throat-microphone | |
| - whisper | |
| - asr | |
| dataset_info: | |
| features: | |
| - name: audio | |
| dtype: | |
| audio: | |
| sampling_rate: 16000 | |
| - name: text | |
| dtype: string | |
| - name: duration | |
| dtype: float32 | |
| splits: | |
| - name: train | |
| num_bytes: 295378780.15999997 | |
| num_examples: 1008 | |
| download_size: 296436154 | |
| dataset_size: 295378780.15999997 | |
| configs: | |
| - config_name: default | |
| data_files: | |
| - split: train | |
| path: data/train-* | |
| # Throat Microphone Dataset | |
| 🤗 **Hugging Face**: [pauljunsukhan/throatmic_codered](https://huggingface.co/datasets/pauljunsukhan/throatmic_codered) | |
| 📦 **GitHub**: [pauljunsukhan/throatmicdata](https://github.com/pauljunsukhan/throatmicdata) | |
| 🚀 **Fine-tuned Model**: | |
| - 🤗 [pauljunsukhan/throatmic_subvocalization_whisper](https://huggingface.co/pauljunsukhan/throatmic_subvocalization_whisper) | |
| - 📦 [pauljunsukhan/whisper_finetuning](https://github.com/pauljunsukhan/whisper_finetuning) | |
| 🔐 **Dataset Access**: | |
| - **Downloading**: The dataset is publicly available. Use `download_dataset.py` (see instructions below) | |
| - **Contributing**: Contributions are very welcome! | |
| 1. Request write access through the Hugging Face dataset page - I'd love to have more contributors! | |
| 2. Once approved, use your Hugging Face token: | |
| ## Dataset Description | |
| - **Homepage:** [GitHub Repository](https://github.com/pauljunsukhan/throatmicdata) | |
| - **Repository:** [Hugging Face Repository](https://huggingface.co/datasets/pauljunsukhan/throatmic_codered) | |
| - **Paper:** N/A | |
| - **Point of Contact:** Paul Han | |
| ### Hardware Setup | |
| - **Throat Microphone**: [CodeRed Assault MOD Tactical Throat Mic Headset](https://coderedheadsets.com/assault-mod-tactical-throat-mic-headset/) | |
| - Uses standard 3.5mm audio jack | |
| - Direct connection to MacBook Air's 3.5mm port | |
| - Note: Many USB-C to 3.5mm microphone adapters are not compatible with throat microphones (CIAA/TRRS) | |
| ### Dataset Summary | |
| A high-quality dataset of throat microphone (laryngophone) recordings specifically designed for fine-tuning Whisper and other speech recognition models. The dataset consists of 605 carefully selected English sentences recorded using a throat microphone, which captures speech through vibrations in the throat rather than air-conducted sound. | |
| This dataset is particularly valuable for: | |
| - Training ASR models for noisy environments | |
| - Adapting speech recognition to throat microphone input | |
| - Developing robust voice activity detection | |
| - Research in alternative speech input methods | |
| ### Dataset Statistics | |
| - **Total Recordings:** 605 | |
| - **Total Duration:** 100.4 minutes | |
| - **Average Duration:** 10.0 seconds (median: 10.0s) | |
| - **Duration Range:** 5.9s - 11.0s | |
| - **Standard Deviation:** 0.3s | |
| Duration Distribution: | |
| - <6s: 0.2% (1 recording) | |
| - 6-8s: 0.7% (4 recordings) | |
| - 8-10s: 98.7% (597 recordings) | |
| - 10-12s: 0.5% (3 recordings) | |
| - 12s or greater: 0% | |
| ### Linguistic Characteristics | |
| **Vocabulary Statistics:** | |
| - Total Words: 7,786 | |
| - Unique Words: 3,113 | |
| - Vocabulary Density: 0.40 | |
| - Average Sentence Length: 13.0 words (range: 9-18) | |
| **Part of Speech Distribution:** | |
| - Nouns: 22.7% | |
| - Proper Nouns: 11.4% | |
| - Prepositions: 11.3% | |
| - Verbs: 11.0% | |
| - Determiners: 9.7% | |
| - Adjectives: 9.2% | |
| - Auxiliaries: 6.4% | |
| - Pronouns: 5.1% | |
| - Coordinating Conjunctions: 4.5% | |
| - Adverbs: 4.5% | |
| - Other: 13.8% | |
| **Complexity Metrics:** | |
| - Average Complexity Score: 8.0 | |
| - Average Tree Depth: 5.5 | |
| - Subordinate Clauses: 10.4% | |
| - Relative Clauses: 8.8% | |
| - Adverbial Clauses: 15.5% | |
| **Complexity Distribution:** | |
| - Simple: 0 sentences | |
| - Moderate: 38 sentences | |
| - Complex: 368 sentences | |
| - Very Complex: 199 sentences | |
| ### Supported Tasks | |
| This dataset is suitable for: | |
| - **Automatic Speech Recognition (ASR)**: Training models to transcribe throat microphone audio | |
| - **Speech-to-Text**: Converting throat microphone recordings to text | |
| - **Voice Activity Detection**: Detecting speech in throat microphone signals | |
| - **Domain Adaptation**: Adapting existing ASR models to throat microphone input | |
| ### Languages | |
| The dataset contains English-language recordings only, with: | |
| - Standard American English pronunciation | |
| - Academic and technical vocabulary | |
| - Complex sentence structures | |
| - High-quality transcriptions | |
| ## Dataset Structure | |
| ### Data Instances | |
| Each instance in the dataset contains: | |
| ```python | |
| { | |
| 'audio': { | |
| 'path': str, # Path to the audio file | |
| 'array': np.array, # The audio signal array | |
| 'sampling_rate': int # 16000 (16kHz) | |
| }, | |
| 'text': str, # The transcription | |
| 'duration': float # Length in seconds | |
| } | |
| ``` | |
| ### Audio Quality | |
| All recordings meet strict quality requirements: | |
| - Sample Rate: 16kHz mono | |
| - Duration: Typically 8-10 seconds (98.7% of recordings) | |
| - Audio Levels: -50dB to 0dB | |
| - Signal-to-Noise Ratio: >10dB | |
| - Silence Ratio: <30% | |
| - Clean, professional recording environment | |
| ### Data Fields | |
| - `audio`: Audio file in WAV format (16kHz mono) | |
| - `text`: String containing the transcription | |
| - `duration`: Float value representing duration in seconds | |
| ### Data Splits | |
| The dataset is provided as a single training split. | |
| ## Dataset Creation | |
| ### Curation Rationale | |
| This dataset was created to address the lack of high-quality throat microphone data for training speech recognition models. Throat microphones are particularly useful in noisy environments as they capture speech directly through throat vibrations. | |
| ### Source Data | |
| #### Initial Data Collection and Normalization | |
| Sentences were carefully selected to ensure: | |
| - Complexity suitable for model training (9-18 words) | |
| - Proper grammar and punctuation | |
| - Mix of statement types | |
| - Natural language patterns | |
| - Varied vocabulary | |
| - Balanced phonetic content | |
| ### Annotations | |
| The annotations (transcriptions) are the original sentences used for recording, ensuring 100% accuracy. | |
| ## Considerations for Using the Data | |
| ### Social Impact of Dataset | |
| This dataset can help improve speech recognition in: | |
| - High-noise environments | |
| - Military and emergency services communications | |
| - Industrial settings | |
| - Assistive technology for voice disorders | |
| ### Discussion of Biases | |
| The dataset: | |
| - Contains only English language | |
| - Uses standard English pronunciation | |
| - May not represent all accents or dialects | |
| - Recorded by a limited number of speakers | |
| ### Other Known Limitations | |
| - Limited to throat microphone recordings | |
| - May not generalize well to regular microphone input | |
| - Optimized for 8-10 second utterances | |
| ## Additional Information | |
| ### Dataset Curators | |
| This dataset was curated by Paul Han | |
| ### Licensing Information | |
| This dataset is released under the MIT License. | |
| ### Citation Information | |
| If you use this dataset, please cite: | |
| ``` | |
| @misc{throatmic_dataset, | |
| title={Throat Microphone Dataset for Speech Recognition}, | |
| author={Han, Paul}, | |
| year={2024}, | |
| publisher={Hugging Face}, | |
| howpublished={\url{https://huggingface.co/datasets/pauljunsukhan/throatmic_codered}} | |
| } | |
| ``` | |