README.md

---
license: cc-by-4.0
language:
- fa
task_categories:
- text-to-speech
- automatic-speech-recognition
- audio-to-audio
- audio-classification
- text-to-audio
- voice-activity-detection
tags:
- TTS
- farsi
- yodas
- quality
pretty_name: YodaLingua
dataset_info:
  features:
  - name: __key__
    dtype: string
  - name: mp3
    dtype:
      audio:
        sampling_rate: 24000
  - name: text
    dtype: string
  - name: language
    dtype: string
  - name: speaker_id
    dtype: string
  - name: dnsmos
    dtype: float64
  splits:
  - name: train
    num_bytes: 680427539.778
    num_examples: 14586
  download_size: 636505858
  dataset_size: 680427539.778
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---

# YodaLingua-Farsi

**YodaLingua** is a high-quality speech dataset designed for training text-to-speech (TTS) systems, ASR models, and any application requiring clean, well-aligned audio–text pairs.  
This release contains the **Farsi** portion of the multilingual YodaLingua collection.

## 🧾 Dataset Overview

| Property | Value |
|---------|-------|
| **Total clips** | 14,586 audio–transcription pairs |
| **Total duration** | 44 hours |
| **Speakers** | 504 distinct speakers |
| **Audio format** | MP3 • mono • 24 kHz • 16-bit |
| **License** | Permissive — commercial use allowed |

All audio clips are noise-reduced, normalized, and matched with accurate transcriptions.

---

## Data Fields

Each entry in the dataset contains the following fields:

| Field | Description |
|------|------------|
| `__key__` | Unique identifier for each sample. |
| `audio` | Path to the audio file associated with the sample (MP3 format). |
| `text` | Ground-truth transcription of the audio segment. |
| `language` | Language code following ISO 639 standards. |
| `speaker_id` | Unique identifier assigned to each speaker. Multiple audio can share the same speaker ID. |
| `dnsmos` | DNSMOS P.835 Overall (OVRL) score estimating perceptual speech quality; higher values indicate cleaner and more intelligible audio. |

---

## 🌍 Multilingual Versions

Other languages are available in the YodaLingua multilingual collection:  
👉 https://huggingface.co/collections/Thomcles/yodalingua

---

We apply a multi-stage pipeline to ensure maximum data quality:

### **1. Standardization**
- Convert to WAV  
- Mono channel  
- Resample to **24 kHz**  
- **16-bit** sample width  
- Normalize to **–20 dBFS** (with volume correction between –3 and +3 dB)

### **2. Noise Reduction**
Advanced denoising applied to improve clarity and remove background artifacts.

### **3. Speaker Diarization**
Segment long recordings by speaker to improve diversity and ensure speaker-consistent utterances.

### **4. Voice Activity Detection (VAD)**
Merge consecutive VAD segments from the same speaker into clean utterances of **3–30 s**.

### **5. Transcription**
State-of-the-art ASR models produce accurate text transcripts.

### **6. Quality Filtering**
Clips are filtered using **DNSMOS P.835 OVRL**; only samples with a score **> 3.0** are retained.


## 📚 Loading the Dataset

```python
from datasets import load_dataset

ds = load_dataset("Thomcles/YodaLingua-Farsi")
```

## Phoneme distribution (as produced by the G2P model)

The following table shows the relative frequency of G2P-generated phoneme units.
These units include vowels, consonants, and G2P-specific markers (e.g., length ː, aspiration ʰ).

This is not intended as a phonological analysis of Farsi, but as an objective indicator of the phonetic diversity and coverage of the dataset for speech-generation tasks.

| Symbol | Frequency |
|--------|-----------|
| ː | 13.94% |
| æ | 10.49% |
| ʰ | 6.25% |
| i | 5.90% |
| ɒ | 5.85% |
| n | 4.95% |
| h | 4.77% |
| d | 4.38% |
| e | 4.25% |
| ɾ | 4.20% |
| m | 4.06% |
| t | 3.53% |
| ʔ | 3.09% |
| b | 2.86% |
| v | 2.31% |
| k | 2.29% |
| u | 2.14% |
| o | 2.10% |
| s | 2.10% |
| ʃ | 1.82% |
| j | 1.78% |
| l | 1.73% |
| z | 1.32% |
| ɡ | 0.78% |
| x | 0.77% |
| f | 0.68% |
| q | 0.64% |
| ʒ | 0.60% |
| p | 0.42% |

---

## Contact

e-mail : [[email protected]]([email protected])