---
license: cc-by-nc-sa-4.0
task_categories:
- text-to-speech
language:
- az
tags:
- azerbaijani
- news
- tts
- texttospeech
- azerbaijani_audio
pretty_name: Azerbaijani News Dataset
size_categories:
- 1K<n<10K
configs:
- config_name: default
  data_files:
  - split: mixed
    path: data/mixed-*
  - split: single
    path: data/single-*
  - split: double
    path: data/double-*
  - split: triple
    path: data/triple-*
dataset_info:
  features:
  - name: audio
    dtype: audio
  - name: text
    dtype: string
  splits:
  - name: mixed
    num_bytes: 6909112496.125
    num_examples: 15225
  - name: single
    num_bytes: 2336567620.12
    num_examples: 8680
  - name: double
    num_bytes: 1875503911.936
    num_examples: 3512
  - name: triple
    num_bytes: 1512994302.324
    num_examples: 1894
  download_size: 12217911129
  dataset_size: 12634178330.505
---


# Audio Dataset Designed for Speech Synthesis in Azerbaijani

Speech Synthesis requires special attention as generated sound should feel natural and grammatically correct. News sessions in TV are the most accurate speech in which it is possible to find voice of the same person or voice of the similar kind. Moreover, mostly content of the speech exists in this type of information, therefore it is easier and faster to validate.

Baku Higher Oil School AI R&D Center (BHOSAI) is presenting the dataset based on news speech with its content. Dataset has been collected from ens.az news agency website. Considering that each data is longer that 1 minute and contains more than 4 5 sentences, we decided to process all these voices and create sentence based speech. We used Whisper model from OpenAI locally and generated transcriptions from voices. Generated transcriptions have been compared with real content and most similar content has been selected as transcription of that sentence.

During model Development (SARA: https://huggingface.co/BHOSAI/SARA_TTS) we have noticed that the end of sentences and punctuations are not generated well. Therefore we decided to create combinations according to the original form of the audio. So of the original audio is 1 minute, several sub-audios are generated and each of them becomes either 2-3 sentences. By this way we increase the robustness of the model in punctuations. Therefore there are split types in this dataset in which you can understand that they are separated according to the combination of the sentences  

![image/png](/static-proxy?url=https%3A%2F%2Fcdn-uploads.huggingface.co%2Fproduction%2Fuploads%2F642bb2d07f152f6e72b8ad2a%2FmONfh80_LEY0SqPTFfe6F.png)

Dataset consist of 4 splits:

- Mixed:  15200 samples containing combination of sentences from 1 to 4. It  means that you can find audio containing 1 sentence or 2 3 4 consecutive sentences
- Single: 8680 samples, each audio is 1 sentence
- Double: 3510 samples, each audio is 2 sentences
- Triple: 1590 samples, each audio is 3 sentences