| --- |
| license: cc-by-4.0 |
| configs: |
| - config_name: default |
| data_files: |
| - split: train |
| path: data/train-* |
| dataset_info: |
| features: |
| - name: audio |
| dtype: audio |
| splits: |
| - name: train |
| num_bytes: 515507464 |
| num_examples: 10 |
| download_size: 463327121 |
| dataset_size: 515507464 |
| --- |
| **Dataset Description:** |
|
|
| This dataset is a **large-scale collection of raw Bengali podcast audio**, specifically designed to support the development and pretraining of speech and language models. |
|
|
| It captures real-world interactions across diverse topics and formats. The dataset preserves natural speech patterns, speaker variability, and authentic podcast environments, making it highly valuable for building robust, scalable, and production-ready AI systems. |
| Additionally, this dataset can be used in data pipelines for **Supervised Fine-Tuning (SFT) and Reinforcement Learning with Human Feedback (RLHF) workflows**. |
|
|
| **Key Use Cases** |
|
|
| -Pretraining Automatic Speech Recognition (ASR) systems |
| -Speech-to-Text (STT) systems |
| -Self-supervised learning (SSL) for speech models |
| -Large Language Models (LLMs) with audio understanding capabilities |
| -Speech representation learning |
| -Noise-robust and real-world voice applications |
| |
| **Dataset Specification** |
|
|
| -Language: Bengali |
| -Type: Raw, unprocessed podcast audio, Single Channel |
| -Speech Style: Natural, conversational, unscripted |
| -Audio Conditions: Real-world environments (including noise and variability) |
| -Domains: discussions, storytelling, interviews, etc. |
| -Format: .wav, .mp3, .ogg, etc. |
| -Sampling Rate: 8000 Hz |
| -Duration: 7798 hours |
| |
| **Value of Single Channel Dataset** |
|
|
| -Training models that can handle real-world conversational complexity |
| -Improved performance in noisy and uncontrolled environments |
| -Development of accurate speaker diarization systems |
| -Better generalization across accents, tones, and speaking styles |
| -Flexible preprocessing and custom annotation pipelines tailored to specific business needs |
| |
| |
| |
| |
| **Audio Quality Analysis** |
| |
|
|
| Signal Quality Analysis (Signal QA) |
|
|
| To ensure robust signal-level integrity and consistency, the dataset was evaluated using multiple acoustic and signal-processing metrics. |
|
|
| | Metric | Value | Interpretation | |
| | ------------------------- | ---------- | ------------------------------------------------------------------------------- | |
| | **Average SNR (dB)** | **50.03** | High signal-to-noise ratio indicating clean audio with minimal background noise | |
| | **Average RMS Energy** | **0.089** | Stable signal energy level, suitable for speech processing tasks | |
| | **Silence Ratio** | **0.448** | reflects natural conversational pauses | |
| | **Clipping Ratio** | **0.0** | No clipping detected, ensuring distortion-free audio | |
| | **Loudness (LUFS)** | **-22.12** | Well-balanced loudness within acceptable range for speech datasets | |
| | **Overall Quality Score** | **70.83** | Good signal quality, appropriate for training and evaluation pipelines | |
|
|
|
|
|
|
| DNSMOS Evaluation |
|
|
| To ensure production-level reliability, the dataset was evaluated using DNSMOS (Deep Noise Suppression Mean Opinion Score) out of 5 is. |
|
|
| | Metric | Score | Interpretation | |
| | ---------------------- | -------- | --------------------------------------------------------- | |
| | Speech Quality (SIG) | **3.89** | Clear and intelligible conversational speech | |
| | Background Noise (BAK) | **4.01** | Strong noise suppression with stable acoustic clarity | |
| | Overall MOS (OVR) | **3.81** | High-quality real-world audio suitable for model training | |
|
|
|
|
| **SQUIM-Based Audio Quality Analysis** |
|
|
| To further assess perceptual and signal characteristics, the dataset was evaluated using SQUIM-based metrics. |
|
|
|
|
| | Metric | Value | Interpretation | |
| | ---------------------------- | --------- | --------------------------------------------------------------------------- | |
| | **Average Energy** | **0.003** | Low energy level, indicating controlled signal amplitude without distortion | |
| | **Spectral Flatness** | **0.052** | Low flatness suggests speech-dominant signal (not noise-like) | |
| | **Zero Crossing Rate (ZCR)** | **0.062** | Low ZCR, consistent with voiced speech and minimal high-frequency noise | |
| | **Dynamic Range** | **1.683** | Moderate variation in amplitude, capturing natural speech dynamics | |
| | **SI-SDR Proxy** | **15.0** | Good signal-to-distortion ratio, indicating clear and well-separated speech | |
| | **SQUIM Score** | **62.59** | Solid perceptual quality, suitable for real-world speech applications | |
|
|
|
|
| Key Insight |
|
|
| The dataset maintains strong acoustic quality despite real-world conditions, making it suitable for production-grade AI systems, LLM pipelines, and speech understanding models. |
|
|
|
|
| **Dataset Validation via End-to-End Model Training** |
|
|
| To validate dataset effectiveness, a complete speech-to-NLP training pipeline was built and executed using InfoBay.AI Audio dataset |
|
|
| **Full Pipeline** |
|
|
| Raw podcast audio |
| → OpenAI Whisper transcription |
| → Sentiment labeling |
| → DistilBERT training (from scratch) |
| → 3-class sentiment classification |
|
|
| Validation Insight |
|
|
| This end-to-end workflow demonstrates that the dataset is not only large-scale but also self-sufficient for training downstream AI models without reliance on external pretrained datasets. |
|
|
|
|
| **Sentiment Classification Task** |
|
|
| The dataset supports supervised learning for sentiment understanding across three classes: |
|
|
| Negative (Class 0) |
| Neutral (Class 1) |
| Positive (Class 2) |
| |
| The dataset contains naturally occurring emotional and contextual variation, making it highly suitable for: |
|
|
| RLHF preference modeling |
| Emotion-aware conversational agents |
| Human-aligned response generation systems |
| |
|
|
|
|
| **Model Performance (From-Scratch Training) From our Dataset** |
|
|
| A DistilBERT-based model trained from scratch achieved strong performance on this dataset: |
| Accuracy: ~98% |
| Macro F1-score: ~0.98 |
| Weighted F1-score: ~0.99 |
| |
| Classification Report |
|
|
| | Class | Sentiment | Precision | Recall | F1-score | Support | |
| | ----- | --------- | --------- | ------ | -------- | ------- | |
| | 0 | Negative | 0.97 | 0.96 | 0.96 | 1,128 | |
| | 1 | Neutral | 0.99 | 0.99 | 0.99 | 7,865 | |
| | 2 | Positive | 0.98 | 0.98 | 0.98 | 2,658 | |
|
|
|
|
|
|
| **Basic JSON Schema** |
| ```json |
| { |
| "id": "string", |
| "audio_filepath": "string", |
| "duration": "float", |
| "language": "string", |
| "sample_rate": "integer", |
| "format": "string", |
| "num_speakers": "integer", |
| "domain": "string", |
| "metadata": { |
| "source": "string", |
| "recording_condition": "string" |
| } |
| } |
| ``` |
|
|
| **Full Dataset Overview** |
|
|
| Total Duration (in hours): 57,568 |
|
|
| This dataset is part of a large multilingual podcast audio collection covering the following languages: |
| Arabic, Bengali, English, Gujarati, Hindi, Kannada, Malayalam, Marathi, Punjabi, Tamil, Telugu, and Urdu. |
|
|
|
|
|
|
| **Data Creation** |
|
|
| Procured through formal agreements and generated in the ordinary course of business. |
|
|
|
|
| **Considerations** |
|
|
| This dataset is provided for research and educational purposes only. It contains only sample data. For access to the full dataset and enterprise licensing options, please visit our website [InfoBay AI](https://infobay.ai/) or contact us directly. |
|
|
| -Ph: (91) 8303174762 |
| -Email: datareq@infobay.ai |
