zuhri025/urdu-turn-detection-audio-v2

🗣️ Urdu Turn Detection (Audio Dataset V2)

This is the official dataset for the model [PuristanLabs1/urdu-turn-v2](https://huggingface.co/PuristanLabs1/urdu-turn-v2), a high precision, low latency system for detecting the end of a conversational turn in Urdu speech.

It contains 11,479 audio clips (balanced between Complete and Incomplete) specifically designed to train robust models for realtime Voice AI applications like "Smart Turn" or "Barge-in" detection.

🚀 How This Dataset Was Compiled

The creation of this dataset involved a multi stage pipeline combining human validated Urdu text, synthetic dialogue generation, and advanced acoustic augmentation.

1. The Core Text Corpus (10,000 Samples)

• Source: 2,825 human validated Urdu sentences (Gold Standard) combined with 7,175 synthetic samples generated by Google Gemini 2.5 Flash Lite.
• Diversity: Covers daily life, news, formal requests, and casual "thinking out loud" scenarios.
• Language: 100% Urdu script (Nastaliq/Arabic), strictly filtered to remove English or Roman Urdu artifacts.

2. Acoustic Realization (VAD Prep)

• We converted the text into audio using high quality Urdu TTS endpoints.
• Each sentence was transformed into a 16kHz mono .wav file.
• Problem: Standard TTS audio ends perfectly. To simulate realworld speech, we needed to simulate "trailing off."

3. Negative Sampling & Truncation

To teach the model what an Incomplete turn sounds like, we implemented a strategic truncation algorithm:

• Full Clips (Label 1): Sentences played to completion with natural closure.
• Truncated Clips (Label 0): Sentences cut off mid-phrase at semantic boundaries (e.g., stopping after "میں..." or "اگر وہ..."). This forces the model to learn the prosodic and phonetic cues of a non-terminal sound.

4. Fixing the "Silence Bias" (V2 Upgrade)

A critical step taken was the injection of Synthetic Silence & Thinking Noise:

• Silence Injection: Added 1,500+ clips of pure, varied energy silence labeled as Incomplete (0).
• Strategic Truncation: Sentences were cut off mid phrase at semantic boundaries to create negative samples.
• Thinking Noise: Injected "hmmm" and "uhh" fillers to ensure the model doesn't trigger a turn end just because the speaker is thinking. Basically Pure silence and "thinking noise" (umms/ahhs) were labeled as Incomplete (0) to prevent false triggers during pauses.
• Padding Masking: All audio is padded to 3.0 seconds, with an explicit attention_mask generated to teach the model to ignore non-voice segments.
• Gemini 2.5 Augmentation: Synthetic dialogues were generated to broaden the linguistic variety beyond standard text corpora.

📊 Dataset Statistics

Feature	Details
Total Clips	11,479
Sampling Rate	16,000 Hz
Format	Mono WAV / PCM_16(WebDataset TAR Shards)
Labels	0: INCOMPLETE, 1: COMPLETE
Avg Duration	2.5 seconds
Balance	57% Complete (6,539) / 43% Incomplete (4,940)

📁 Repository Structure

The dataset is stored in WebDataset format (TAR shards) for optimal streaming:

• train-0000.tar to train-0005.tar: Audio shards.
• metadata.csv: Combined metadata.
• Each sample inside a shard contains:
  • XXXX.wav: Audio clip.
  • XXXX.txt: Urdu transcription.
  • XXXX.cls: Label (0 or 1).

🛠️ Installation & Usage

⚙️ Dependencies (Crucial for Windows)

The datasets library uses torchcodec for audio decoding. Windows users often face issues because torchcodec depends on FFmpeg DLLs.

1. Install Python Packages:

pip install datasets librosa torchcodec
or

!pip install -q datasets[audio] librosa Transformers

2. Install FFmpeg (Required for Windows):

If you see ImportError: To support decoding audio data, please install 'torchcodec' or invalid header errors:

• Using Chocolatey: choco install ffmpeg-full
• Manual: Download the "full-shared" build from gyan.dev, extract it, and add the bin folder to your System PATH.

---

🚀 Loading the Dataset

Option 1: Streaming Mode (Recommended)

This is the fastest way to start. It fetches audio on the fly without downloading the whole 1GB archive.

from datasets import load_dataset
import io
import librosa

# Load in streaming mode
ds = load_dataset("PuristanLabs1/urdu-turn-detection-audio-v2", streaming=True, split="train")

for sample in ds.take(5):
    # Access audio data
    audio_array = sample["wav"]["array"]
    sampling_rate = sample["wav"]["sampling_rate"]
    
    # Access metadata
    text = sample["txt"]
    label = int(sample["cls"]) # 0: Incomplete, 1: Complete
    
    print(f"Text: {text} | Label: {label}")

### 🔊 Playing the Audio

Once you have the `audio_array` from the sample:

#### **On Google Colab:**
```python
import IPython.display as ipd
ipd.display(ipd.Audio(audio_array, rate=sampling_rate))

On Local Machine (Laptop):

• Windows: pip install sounddevice
• Linux (Ubuntu/Debian): sudo apt-get install libportaudio2 then pip install sounddevice

import sounddevice as sd
sd.play(audio_array, sampling_rate)
sd.wait() # Wait for audio to finish

Option 2: Full Download

Download the entire dataset for offline training.

from datasets import load_dataset

ds = load_dataset("PuristanLabs1/urdu-turn-detection-audio-v2")
print(f"Total samples: {len(ds['train'])}")

---

🤝 Acknowledgments & Credits

• Curated By: PuristanLabs.
• Base Audio: Custom TTS on our PuristanLabs1/Urdu-Turn-Detection-10k & Common Voice 13 (Urdu Subset)
• Augmentation: Synthetic generation via Google Gemini 2.5.
• Core Engine: Optimized for the urdu-turn-detector library.

---

License: MIT. Free for research and commercial use.