README.md

---
dataset_info:
- config_name: long_digit
  features:
  - name: file
    dtype: string
  - name: audio
    dtype: audio
  - name: instruction
    dtype: string
  - name: audio_transcription
    dtype: string
  - name: answer
    dtype: string
  splits:
  - name: kazakh
    num_bytes: 32786850
    num_examples: 172
  download_size: 32758406
  dataset_size: 32786850
- config_name: multi_step_reasoning
  features:
  - name: file
    dtype: string
  - name: audio
    dtype: audio
  - name: instruction
    dtype: string
  - name: audio_transcription
    dtype: string
  - name: answer
    dtype: string
  splits:
  - name: kazakh
    num_bytes: 1094374976
    num_examples: 1392
  download_size: 1093760168
  dataset_size: 1094374976
- config_name: short_digit
  features:
  - name: file
    dtype: string
  - name: audio
    dtype: audio
  - name: instruction
    dtype: string
  - name: audio_transcription
    dtype: string
  - name: answer
    dtype: string
  splits:
  - name: kazakh
    num_bytes: 17245162
    num_examples: 100
  download_size: 17231695
  dataset_size: 17245162
- config_name: single_step_reasoning
  features:
  - name: file
    dtype: string
  - name: audio
    dtype: audio
  - name: instruction
    dtype: string
  - name: audio_transcription
    dtype: string
  - name: answer
    dtype: string
  splits:
  - name: kazakh
    num_bytes: 276560211
    num_examples: 592
  download_size: 276402538
  dataset_size: 276560211
configs:
- config_name: long_digit
  data_files:
  - split: kazakh
    path: long_digit/kazakh-*
- config_name: multi_step_reasoning
  data_files:
  - split: kazakh
    path: multi_step_reasoning/kazakh-*
- config_name: short_digit
  data_files:
  - split: kazakh
    path: short_digit/kazakh-*
- config_name: single_step_reasoning
  data_files:
  - split: kazakh
    path: single_step_reasoning/kazakh-*
---

# SpokenMQA — Kazakh

This dataset is a machine-translated Kazakh adaptation of the [SpokenMQA benchmark](https://huggingface.co/datasets/amao0o0/spoken-mqa) for evaluating spoken mathematical reasoning in speech-based and audio-language models.

The original SpokenMQA benchmark probes whether speech-based models can solve math problems delivered through speech, ranging from simple digit recognition (short and long digits) to single-step and multi-step mathematical reasoning. This release localizes the benchmark into Kazakh, enabling evaluation of multilingual speech models on spoken mathematical reasoning beyond English.

The textual fields of the original benchmark were machine-translated into Kazakh, and the audio was re-synthesized in Kazakh from the translated transcripts using [MangiSozTTS](https://mangisoz.nu.edu.kz/).

## Dataset structure

The benchmark covers four tracks of increasing difficulty:

| Subset                   | Purpose                                                                  |
|--------------------------|--------------------------------------------------------------------------|
| `short_digit`            | Recognition of short numerical sequences spoken aloud                    |
| `long_digit`             | Recognition of longer numerical sequences spoken aloud                   |
| `single_step_reasoning`  | Mathematical problems requiring a single arithmetic operation            |
| `multi_step_reasoning`   | Mathematical problems requiring multi-step reasoning over spoken input   |