README.md

metadata

language:
  - pa
license: apache-2.0
task_categories:
  - text-to-speech
  - automatic-speech-recognition
  - audio-classification
multilinguality:
  - monolingual
size_categories:
  - n<1K
tags:
  - punjabi
  - gurmukhi
  - punjabi-voice
  - punjabi-audio
  - synthetic-speech
  - tts
  - pure-punjabi
  - speech-recognition
pretty_name: Punjabi (Gurmukhi) Synthetic Multi-Generation Voice Corpus

🎙️ Punjabi (Gurmukhi) Synthetic Multi-Generation Voice Corpus

☬ ਸ਼ੁੱਧ ਗੁਰਮੁਖੀ ਮਹਾਨ ਕੋਸ਼ ਸਿੰਥੈਟਿਕ ਵੌਇਸ ਡਾਟਾਸੈੱਟ

⚠️ Corrected 2026-09-02. The original README described this as "5 distinct acoustic age & gender profiles" of "Studio Master" recordings. That was false. This is 100% synthetic, machine-generated speech — not a multi-speaker human recording. See "How the audio was made" below.

A Punjabi (Gurmukhi) synthetic speech corpus built from two underlying text-to-speech engines, covering everyday conversational, health-coach, and smart-device-prompt style sentences in pure Gurmukhi vocabulary (minimizing Hindi loan words).

---

🔊 How the audio was actually made

• Base voice A (female): Google's cloud TTS via gTTS, Punjabi (pa).
• Base voice B (male): Meta's facebook/mms-tts-pan VITS model.
• "Child", "Elder Male / Baba Ji", "Elder Female / Bibi Ji" variants are NOT separate speakers or separate TTS models. They are the same base gTTS voice, pitch/tempo-shifted with ffmpeg audio filters (e.g. asetrate=16000*1.28,atempo=1/1.28 for the "child" variant, asetrate=16000*0.82,atempo=1/0.82 for "elder male"). This is a DSP trick applied to one voice, not a real distinct demographic speaker.

If you need genuinely distinct human speakers for voice-model training, this dataset is not that — treat it as labeled synthetic TTS output only.

---

📊 Dataset Inventory

• Total Audio Clips: synthetic WAV files (16kHz Mono PCM), generated via gTTS + MMS-TTS as described above.
• Metadata: metadata.csv with Gurmukhi text, category, and audio durations.

---

🚀 Quick Usage in Python

from datasets import load_dataset

dataset = load_dataset("Nam-toon-studio/Punjabi-Studio-Voice-Corpus")
print(dataset['train'][0])