---
language:
- en
- de
- tr
- vi
- zh
task_categories:
- automatic-speech-recognition
tags:
- code-switching
- multilingual
- long-form
- asr
source_datasets:
- goodpiku/muscat-eval
size_categories:
- n<1K
dataset_info:
  features:
  - name: id
    dtype: string
  - name: audio
    dtype:
      audio:
        sampling_rate: 16000
  - name: transcription
    dtype: string
  - name: transcription_tagged
    dtype: string
  - name: duration_sec
    dtype: float64
  - name: languages
    list: string
  - name: device
    dtype: string
  - name: conv_lang
    dtype: string
  - name: recording
    dtype: int64
  splits:
  - name: test
    num_bytes: 369818230
    num_examples: 18
  download_size: 369676385
  dataset_size: 369818230
configs:
- config_name: default
  data_files:
  - split: test
    path: data/test-*
---

# MUSCAT — Merged Long-Form Samples

This dataset is a **merged, long-form** reformatting of
[`goodpiku/muscat-eval`](https://huggingface.co/datasets/goodpiku/muscat-eval)
(MUSCAT: A Multi-Device Dataset for Code-Switching ASR and Segmentation
Evaluation).

The original MUSCAT release stores each conversation as many short,
single-language segments. Here those segments are **concatenated back into one
continuous recording per conversation**, so each row is a single long-form
code-switching audio with inline language/timing markers. The layout mirrors
[`BrunoHays/fleurs_code_switching_test`](https://huggingface.co/datasets/BrunoHays/fleurs_code_switching_test)
to enable code-switching ASR evaluation with consistent tooling.

## How it was built

* **Source:** `goodpiku/muscat-eval`, `manual` segmentation only (the
  configuration carrying ground-truth text).
* **Grouping:** segments are grouped by `(device, recording)` and ordered by
  their segment index, then concatenated.
* **Audio:** each segment is decoded, downmixed to **mono** and resampled to
  **16 kHz**, then concatenated into a single waveform.
* **Timestamps:** the `<start:..>` / `<end:..>` markers are recomputed from the
  cumulative duration of each merged segment (not the original file
  timestamps), so they reflect the merged audio timeline.
* **Languages:** taken from the per-segment `lid` field (falling back to the
  language word encoded in the source filename when `lid` is empty).

Each of the 6 conversations is recorded by 3 devices (`owl`, `aria`, `pi`),
yielding **18 merged samples**.

## Columns

* `id`: unique id for the merged sample.
* `audio`: concatenated waveform (16 kHz mono).
* `transcription`: plain concatenation of the segment transcripts.
* `transcription_tagged`: transcript with inline markers per segment, formatted
  as `<lang><start:SS.ss>text<end:SS.ss>`, e.g.

  ```
  <de><start:0.00>Hallo, Yining. ...<end:5.23><en><start:5.23>Yes, I find it really interesting...<end:12.52>
  ```

* `duration_sec`: total duration of the merged sample in seconds.
* `languages`: sorted unique languages present in the sample.
* `device`: recording device (`owl`, `aria`, `pi`).
* `conv_lang`: primary language pair of the conversation (`en-de`, `en-tr`, `en-vi`, `en-zh`).
* `recording`: source recording/conversation index.

## Usage

```python
from datasets import load_dataset

ds = load_dataset("BrunoHays/muscat-merged-samples", split="test")
print(ds[0]["transcription_tagged"][:200])
```

## Limitations

* Audio is downmixed to mono 16 kHz; multi-device spatial information from the
  original recordings is not preserved.
* Segment joins are concatenative, so prosody and speaker continuity across
  joins are not guaranteed to be natural.

## Citation

This dataset is derived from the **MUSCAT** benchmark. Please cite the original
dataset:

```bibtex
@misc{muscat_eval,
  title        = {MUSCAT: A Multi-Device Dataset for Code-Switching ASR and Segmentation Evaluation},
  howpublished = {Hugging Face dataset \url{https://huggingface.co/datasets/goodpiku/muscat-eval}},
  note         = {Source dataset merged into long-form samples in BrunoHays/muscat-merged-samples}
}
```