Datasets:

typhoon-ai
/

TVSpeech

TVSpeech is a Thai speech recognition benchmark dataset specifically designed as a Robustness Track for evaluating ASR models on real-world, in-the-wild Thai audio. The dataset consists of 570 utterances (3.75 hours) curated from diverse public media channels on YouTube under the Creative Commons Attribution (CC-BY) license, representing challenging acoustic and semantic complexity found in natural speech.

Dataset Overview

Language: Thai
Domain: In-the-wild video speech from public media (YouTube)
Samples: 570 utterances
Total Duration: 3.75 hours
Sampling Rate: 16 kHz
Task: ASR Robustness Evaluation (Robustness Track)
License: Creative Commons Attribution (CC-BY)
Source: Diverse public media channels on YouTube

Dataset Structure

The dataset contains a single test split intended for evaluation purposes only.

Data Fields

audio_id (string): Unique identifier for each audio sample
path (string): Original file path reference
audio (Audio): Audio file sampled at 16kHz
sentence (string): Manual transcription of the audio

Data Statistics

Total samples: 570
Average transcription length: 279.72 characters
Min/Max transcription length: 12 / 505 characters
Average word count: 14.04 words
Min/Max word count: 1 / 36 words

Example

from datasets import load_dataset

dataset = load_dataset("typhoon-ai/TVSpeech")

# Access the test split
test_data = dataset['test']

# View a sample
print(test_data[0])
# {
#   'audio_id': '5c6hg2Vn43M_24416_53728',
#   'path': '...',
#   'audio': {'path': '5c6hg2Vn43M_24416_53728.wav', 'sampling_rate': 16000, 'array': [...]},
#   'sentence': 'ของการเมือง มันมีปัจจัยสำคัญอยู่สามสี่อย่าง...'
# }

Dataset Characteristics

Robustness Track Design

TVSpeech is specifically designed to measure ASR performance in "in-the-wild" conditions that challenge both acoustic and semantic understanding.

Manual Curation Strategy

Unlike random sampling approaches, this test set was manually curated to maximize both acoustic and semantic complexity:

High Lexical Density: Annotators prioritized segments containing domain-specific terminology, proper names, and technical jargon
Diverse Content Categories: Includes Finance, Technology, and Variety Vlogs
Low-Frequency Entity Focus: Tests the model's ability to accurately resolve uncommon terms and proper nouns
Minimal Language Model Dependence: Designed to evaluate transcription accuracy without excessive reliance on language model hallucination

Key Evaluation Aspects

TVSpeech evaluates ASR models on:

Acoustic Robustness: Performance on noisy, real-world audio conditions typical of video content
Semantic Complexity: Handling of domain-specific terminology and technical vocabulary
Entity Resolution: Accurate transcription of proper names and low-frequency terms
Lexical Coverage: Ability to transcribe specialized content across multiple domains

This makes TVSpeech particularly valuable for:

Benchmarking ASR robustness in challenging real-world scenarios
Evaluating model performance on specialized and technical Thai content
Testing transcription accuracy on low-frequency vocabulary without language model over-correction
Comparing model performance across diverse content categories

Usage

Loading the Dataset

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("typhoon-ai/TVSpeech")

# Access test samples
for sample in dataset['test']:
    audio = sample['audio']
    transcription = sample['sentence']
    # Your evaluation code here

Dataset Creation

Source and Curation

TVSpeech was curated from diverse public media channels on YouTube under the Creative Commons Attribution (CC-BY) license. Unlike random sampling approaches, the dataset was manually curated to maximize both acoustic and semantic complexity.

Selection Criteria

Annotators prioritized segments with high lexical density, specifically selecting clips containing:

Domain-specific terminology across Finance, Technology, and Variety Vlogs
Proper names and entity references
Technical jargon and specialized vocabulary
Low-frequency terms that challenge language model predictions

This selection strategy ensures the dataset evaluates ASR models not just on background noise robustness, but on their ability to accurately resolve low-frequency entities without relying on excessive language model hallucination or over-correction.

Processing

All audio samples were:

Sourced from YouTube public media channels (CC-BY licensed)
Manually transcribed by native Thai speakers
Normalized to 16 kHz sampling rate
Validated for transcription quality and accuracy
Processed using the Typhoon data pipeline for consistent formatting

Limitations

Test-only dataset: Contains only 570 samples intended for evaluation, not for training
Domain specificity: Focuses on conversational speech, may not represent all Thai speech domains
Size: Relatively small dataset suitable for benchmarking but not comprehensive coverage
Regional variations: May not fully represent all Thai dialects and accents

Citation

If you use this dataset in your research or application, please cite our technical report:

@misc{warit2026typhoonasr,
      title={Typhoon ASR Real-time: FastConformer-Transducer for Thai Automatic Speech Recognition}, 
      author={Warit Sirichotedumrong and Adisai Na-Thalang and Potsawee Manakul and Pittawat Taveekitworachai and Sittipong Sripaisarnmongkol and Kunat Pipatanakul},
      year={2026},
      eprint={2601.13044},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={[https://arxiv.org/abs/2601.13044](https://arxiv.org/abs/2601.13044)}, 
}

Contact

For questions or feedback about this dataset, please visit opentyphoon.ai or open an issue on the dataset repository.