README.md

---
license: mit
task_categories:
  - automatic-speech-recognition
language:
  - ar
  - fa
  - de
  - en
tags:
  - code-switching
  - multilingual
  - asr-benchmark
  - arabic
  - persian
  - german
  - benchmark
pretty_name: ASR Code-Switching Benchmark
size_categories:
  - n<1K
---

# ASR Code-Switching Benchmark

A curated benchmark of **1,200 code-switching utterances** (300 per language pair)
for evaluating commercial ASR systems on multilingual speech with intra-sentential
language switching.

## Paper

> **Benchmarking Commercial ASR Systems on Code-Switching Speech: Arabic, Persian, and German**
> [arXiv link](https://arxiv.org/abs/2605.19069)

## Language pairs

| Split | Language pair | Samples | Scripts |
|---|---|---|---|
| `egyptian_arabic_english` | Egyptian Arabic–English | 300 | Arabic + Latin |
| `saudi_arabic_english` | Saudi Arabic (Najdi/Hijazi)–English | 300 | Arabic + Latin |
| `persian_english` | Persian (Farsi)–English | 300 | Perso-Arabic + Latin |
| `german_english` | German–English | 300 | Latin | 

## Loading the dataset

```python
from datasets import load_dataset

# Load a single split
ds = load_dataset("Perle-ai/ASR_Code_Switch", "egyptian_arabic_english")

# Load all splits
ds = load_dataset("Perle-ai/ASR_Code_Switch")
```

## Sample selection methodology

Samples were selected using a two-stage pipeline designed to surface the
hardest code-switching instances from each source corpus:

**Stage 1 — Heuristic filter** scores each transcript on five structural
signals: script mix ratio, token alternation rate, morphological blend
detection, length, and vocabulary diversity (type-token ratio).

**Stage 2 — LLM ensemble** sends candidates to GPT-4o and Gemini 1.5 Pro,
which independently score each transcript across six linguistic dimensions.
The top 300 per language pair by ensemble score are retained.

## Fields

| Field | Type | Description |
|---|---|---|
| `audio` | Audio | MP3 audio file |
| `audio_file` | string | Original filename |
| `transcript` | string | Human-annotated reference transcript |
| `language_pair` | string | Language pair label |
| `iso_codes` | list | BCP-47 codes for the two languages |
| `gender` | string | Speaker gender (MALE / FEMALE) |
| `h_score` | float | Composite heuristic difficulty score (0–10) |
| `h_script_mix` | float | Script mix ratio signal (0–10) |
| `h_alternation` | float | Token alternation rate signal (0–10) |
| `h_morph_blend` | float | Morphological blend detection signal (0–10) |
| `h_length` | float | Length signal (0–10) |
| `h_unique_ratio` | float | Vocabulary diversity / TTR signal (0–10) |
| `ensemble_score` | float | Mean LLM difficulty score (1–10) |
| `gpt_score` | int | GPT-4o overall difficulty score (1–10) |
| `gemini_score` | int | Gemini 1.5 Pro overall difficulty score (1–10) |
| `gpt_summary` | string | GPT-4o free-text difficulty summary |
| `gemini_summary` | string | Gemini 1.5 Pro free-text difficulty summary |
| `gpt_morphological_blending_score` | int | GPT-4o score for morphological blending (1–10) |
| `gpt_switching_density_score` | int | GPT-4o score for switching density (1–10) |
| `gpt_slang_and_register_mix_score` | int | GPT-4o score for slang/register mix (1–10) |
| `gpt_phonological_ambiguity_score` | int | GPT-4o score for phonological ambiguity (1–10) |
| `gpt_named_entity_jargon_density_score` | int | GPT-4o score for named entity/jargon density (1–10) |
| `gpt_script_orthographic_complexity_score` | int | GPT-4o score for script/orthographic complexity (1–10) |
| `gpt_hard_tokens` | string | JSON list of hard tokens with ASR risk explanations (GPT-4o) |
| `gemini_*` | various | Gemini 1.5 Pro equivalents of all GPT-4o fields above |
| `disagreement_*` | int | Absolute score difference between GPT-4o and Gemini per dimension |
| `max_dimension_disagreement` | int | Maximum disagreement across all six dimensions |
| `any_high_disagreement` | bool | True if any dimension disagreement exceeds threshold |

## Citation

```bibtex
@misc{abdoli2026benchmarkingcommercialasrsystems,
      title={Benchmarking Commercial ASR Systems on Code-Switching Speech: Arabic, Persian, and German}, 
      author={Sajjad Abdoli and Ghassan Al-Sumaidaee and Clayton W. Taylor and Ahmad and ElShiekh and Ahmed Rashad},
      year={2026},
      eprint={2605.19069},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2605.19069}, 
}
```

## Licence

MIT