---
license: cc-by-4.0
language:
  - id
  - en
task_categories:
  - automatic-speech-recognition
tags:
  - code-switching
  - synthetic-speech
  - low-resource
  - indonesian
  - text-to-speech
pretty_name: Indonesian-English Code-Switching Synthetic Speech
---

# Indonesian–English Code-Switching Synthetic Speech Dataset

Synthetic speech generated for the undergraduate final project *"Handling
Code-Switching in Automatic Speech Recognition for Low-Resource Language
Pairs: An Indonesian–English Case Study"*, School of Electrical Engineering
and Informatics, Institut Teknologi Bandung.

This dataset contains synthetic audio produced from the Indonesian–English
code-switching text corpora released in the companion repository below. It
was used as one of the training data sources for adapting Whisper to
intra-sentential Indonesian–English code-switching.

**Companion repository (text corpora, evaluation set, pseudo-labels, code):**
https://github.com/shulhajws/codeswitch-asr-id-en

---

## How this audio was generated

Audio was synthesized with [Chatterbox](https://github.com/resemble-ai/chatterbox)
(Resemble AI, MIT licence), using the Indonesian fine-tune
[Chatterbox-TTS-Indonesian](https://huggingface.co/grandhigh/Chatterbox-TTS-Indonesian)
by grandhigh, from the sentences in the `text_corpus/` folder of the
companion GitHub repository. Each sentence was synthesized in a single pass
as a whole code-switched utterance.

Generation is deterministic given the same model, voice reference, and
configuration, so this audio can be regenerated from the text corpora rather
than downloaded, if needed.

Every clip generated by Chatterbox embeds an imperceptible PerTh neural
watermark. This does not affect the audio's usability for speech recognition
experiments.

---

## Dataset contents

| Split | Sentences | Description |
|---|---|---|
| Synthetic speech | 20,219 | Audio generated from `indo_eng_codeswitch_text_20000_tagged_phonetic.csv` |

Each audio file corresponds to one row in `manifest_with_transcript.csv`
(included in this repository), matched by `id`. Audio files are organized
into batch subfolders (`batch_0000/`, `batch_0001/`, ... `batch_0040/`) to
stay within Hugging Face's per-directory file limit; the exact path for
each file is given in the `audio_path` column.

**`manifest_with_transcript.csv`** columns:

| Column | Description |
|---|---|
| `id` | Sentence identifier, matches the source text corpus |
| `audio_path` | Path to the corresponding audio file, relative to the repository root |
| `ground_truth` | Clean reference transcription |
| `transcript` | The same sentence with inline `<\|id\|>` / `<\|en\|>` language span tags |
| `tts_input` | The phonetically-respelled text actually given to the TTS model (English words respelled to guide pronunciation, e.g. *review* → *rifyu*) |
| `sample_rate` | Audio sample rate in Hz |

Word-level language annotation beyond the inline tags, and the full set of
text corpora this audio was generated from, are in the companion GitHub
repository, not duplicated here, to avoid the two copies drifting out of
sync.

---

## Intended use

This dataset was created to test whether synthetic code-switched speech
can substitute for real recordings when adapting a pretrained ASR model
under limited labelled data. In the accompanying study, models trained on
this synthetic audio were compared against models trained on real
pseudo-labelled YouTube audio; the real audio consistently gave better
code-switching performance, and this synthetic set is released mainly to
support that comparison and for others exploring the same question on
other language pairs.

It is not recommended as the sole training source for a production ASR
system, given the acoustic gap between synthetic and spontaneous speech
noted in the accompanying report.

---

## Licence

This dataset is released under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
Attribution should reference this dataset and the companion GitHub
repository above.

Chatterbox itself is MIT-licensed; see the
[Chatterbox repository](https://github.com/resemble-ai/chatterbox) for
details on the underlying model.