Tim2190's picture
Upload 2 files
540f00c verified
|
Raw
History Blame Contribute Delete
2.83 kB
metadata
license: cc-by-4.0
language:
  - kk
  - ru
task_categories:
  - automatic-speech-recognition
tags:
  - kazakh
  - russian
  - code-switching
  - speech
  - asr
  - benchmark
pretty_name: Kazakh Code-Switching ASR Benchmark
size_categories:
  - n<1K

Kazakh Code-Switching ASR Benchmark

A benchmark for evaluating ASR systems on natural Kazakh speech that code-switches with Russian — the everyday Kazakh–Russian mixing found in stand-up, interviews and vlogs, not scripted read speech. This is, to our knowledge, the first speech/ASR resource targeting the Kazakh–Russian code-switching pair (existing Kazakh–Russian NLP resources are text-only).

Code, scoring harness and full analysis: https://github.com/Tim2190/Kaz-ASR-codeswitch-benchmark

Dataset structure

31 clips (~3.7 min total, clips ≤ 20 s), sourced from CC BY YouTube material. Each clip carries two independent reference transcripts and four binary phenomenon flags:

Field Description
file_name path to the WAV clip
audio_id clip identifier
transcript_verbatim literal transcript, exactly as spoken (reductions kept)
transcript_normalized_written dictionary word forms
has_contraction fast-speech reductions present
has_dialect_slang dialectal/slang lexis present
has_barbarisms a Russian insertion where a native Kazakh word exists
has_propers proper nouns present
audio_source_link source channel/title (CC BY attribution)

Audio is 48 kHz stereo 16-bit WAV. The two-layer design separates genuine recognition errors from a system's tendency to auto-normalize morphology. Full annotation rules are in annotation_methodology.md in the GitHub repo.

Loading

from datasets import load_dataset
ds = load_dataset("Tim2190/kazakh-codeswitch-asr", split="train")
print(ds[0]["audio"], ds[0]["transcript_verbatim"])

Benchmark results (5 systems, WER / CER, lower is better)

System Type WER CER
Fine-tuned Kazakh Whisper fine-tuned 11.9% 3.8%
Yandex SpeechKit commercial STT 17.1% 6.5%
Gemini 2.5 Flash commercial LLM 29.5% 12.5%
Whisper large-v3 base, zero-shot 42.5% 13.0%
Google Cloud STT commercial STT 64.9% 38.1%

Key finding: commercial services span the entire quality range (Yandex 2nd, Google last), and Russian insertions are the dominant error driver — errors concentrate on the switched-in Russian words. See the GitHub repo for the per-phenomenon breakdown and word-level analysis.

License & attribution

Source audio is derived from Creative Commons CC BY YouTube material; the originating channel/title for every clip is in the audio_source_link field. Annotations and code are MIT-licensed. Created by Timur Seidalin.