README.md

---
license: apache-2.0
task_categories:
- audio-to-audio
tags:
- audio-super-resolution
---

# LJSpeech-1.1 High-Resolution Dataset (48,000 Hz)

This dataset was created using the method described in [HiFi-SR: A Unified Generative Transformer-Convolutional Adversarial Network for High-Fidelity Speech Super-Resolution](https://huggingface.co/papers/2501.10045).

The LJSpeech-1.1 dataset, widely recognized for its utility in text-to-speech (TTS) and other speech processing tasks, has now been enhanced through a cutting-edge speech 
super-resolution algorithm. The original dataset, which featured a sampling rate of 22,050 Hz, has been upscaled to 48,000 Hz using [**ClearerVoice-Studio**](https://github.com/modelscope/ClearerVoice-Studio), providing a high-fidelity version suitable 
for advanced audio processing tasks [1].

**Key Features**

  - High-Resolution Audio: The dataset now offers audio files at a sampling rate of 48,000 Hz, delivering enhanced perceptual quality with richer high-frequency details.
  
  - Original Content Integrity: The original linguistic content and annotation structure remain unchanged, ensuring compatibility with existing workflows.
  
  - Broader Application Scope: Suitable for professional-grade audio synthesis, TTS systems, and other high-quality audio applications.
  
  - Open Source: Freely available for academic and research purposes, fostering innovation in the speech and audio domains.

**Original Dataset**

  - Source: The original LJSpeech-1.1 dataset contains 13,100 audio clips of a single female speaker reading passages from public domain books.
  
  - Duration: Approximately 24 hours of speech data.
  
  - Annotations: Each audio clip is paired with a corresponding text transcript.

**Super-Resolution Processing**

The original 22,050 Hz audio recordings were processed using a state-of-the-art MossFormer2-based speech super-resolution model. This model employs:

  - Advanced Neural Architectures: A combination of transformer-based sequence modeling and convolutional networks.
  
  - Perceptual Optimization: Loss functions designed to preserve the naturalness and clarity of speech.
  
  - High-Frequency Reconstruction: Algorithms specifically tuned to recover lost high-frequency components, ensuring smooth and artifact-free enhancement.

**Output Format**

  - Sampling Rate: 48,000 Hz
  
  - Audio Format: WAV
  
  - Bit Depth: 16-bit
  
  - Channel Configuration: Mono

**Use Cases**

  1. Text-to-Speech (TTS) Synthesis
      ○ Train high-fidelity TTS systems capable of generating human-like speech.
      ○ Enable expressive and emotionally nuanced TTS outputs.
     
  2. Speech Super-Resolution Benchmarking
      ○ Serve as a reference dataset for evaluating speech super-resolution algorithms.
      ○ Provide a standardized benchmark for perceptual quality metrics.
      
  3. Audio Enhancement and Restoration
      ○ Restore low-resolution or degraded speech signals for professional applications.
      ○ Create high-quality voiceovers and narration for multimedia projects.

**File Structure**
  
The dataset retains the original LJSpeech-1.1 structure, ensuring ease of use:

  ```sh
  LJSpeech-1.1-48kHz/
  ├── metadata.csv         # Text transcriptions and audio file mappings
  ├── wavs/               # Directory containing 48,000 Hz WAV files
  └── LICENSE.txt         # License information
  ```
**Licensing**

The LJSpeech-1.1 High-Resolution Dataset is released under the same open license as the original LJSpeech-1.1 dataset (https://keithito.com/LJ-Speech-Dataset/). Users are free to use, modify, and share the dataset for academic and non-commercial purposes, provided proper attribution is given.

[1] Shengkui Zhao, Kun Zhou, Zexu Pan, Yukun Ma, Chong Zhang, Bin Ma, "[HiFi-SR: A Unified Generative Transformer-Convolutional Adversarial Network for High-Fidelity Speech Super-Resolution](https://arxiv.org/abs/2501.10045)", ICASSP 2025.