---
tags:
  - audio

configs:
- config_name: test
  data_files: "data/test.parquet"
- config_name: small_test
  data_files: "data/small_test.parquet"
---
### About dataset
It is a dataset of multispeaker speech with noise  
each sample is at most 30 seconds 

### Loading script
```
>>> data_files = {"train": "data/<your_subset>.parquet"}
>>> data = load_dataset("Zarakun/speakers_ua_test", data_files=data_files)
>>> data
DatasetDict({
    test: Dataset({
        features: ['num_speakers', 'utter', 'audio'],
        num_rows: <some_number>
    })
})
```

### Dataset structure
Every example has the following:  
**num_speakers** - the number of speakers    
**utter** - list of utterences data  
**audio** - the waveform of the audio  

Each entry in the **utter** is a dict and has the following structure:  
**start** - the starting position in **audio** of the speaker audio  
**end** - the ending position in **audio** of the speaker audio  
**file_id** - the identificator of the speaker  
**sentence** - the transcription  
**rate** - has to be the same across all examples in the dataset, if not something bad happened