Hugging Face's logo

Datasets:
yigagilbert
/
synthetic-parallel-external

Tasks:
Automatic Speech Recognition
Translation
Modalities:
Audio
Text
Formats:
parquet
Languages:
English
Ganda
Size:
100K - 1M
Libraries:
Datasets
Dask
Polars
License:
Dataset card Data Studio Files
xet
 Community
synthetic-parallel-external / README.md
yigagilbert's picture
yigagilbert
Add dataset card
2ba5905 verified
preview code
|
raw
history blame contribute delete
2.48 kB
metadata 
language:
  - en
  - lug
license: cc-by-4.0
task_categories:
  - automatic-speech-recognition
  - translation
pretty_name: Synthetic Parallel EN-LG (external)

Synthetic Parallel EN↔LG — external

Voice-controlled synthetic parallel speech dataset for Luganda-English speech-to-speech translation, generated by the Hibiki-Zero fine-tuning pipeline.

Generation

Component Model
Translation Sunbird/translate-nllb-3.3b-salt
TTS Sunbird/orpheus-3b-tts-multilingual

English speakers: salt_eng_0001, salt_eng_0002, salt_eng_0003

Luganda speakers: salt_lug_0001, waxal_lug_0001, waxal_lug_0002, waxal_lug_0003, waxal_lug_0004, waxal_lug_0005, waxal_lug_0007

Speaker sampling: N ∈ {1 (70%), 2 (20%), 3 (10%)} draws per text pair.

Source datasets (transcripts only — source audio discarded)

  • evie-8/afrispeech-200 (subset: default, lang: eng)
  • yigagilbert/combined-audio-eng (subset: default, lang: eng)
  • yigagilbert/cv24-lug-eng (subset: default, lang: eng)

Post-processing

  • Silero / energy VAD silence trim on every clip
  • Duration filter: Luganda 1.5–30.0 s, English 1.0–30.0 s
  • Duration ratio filter: 0.3–4.0
  • Speech-activity filter: ≥ 0.35
  • Sample rate: 24000 Hz (mono, float32 → PCM-16 WAV)

Schema

Column Type Description
audio_lug Audio Luganda synthesized clip
audio_eng Audio English synthesized clip
text_lug string Luganda transcript
text_eng string English transcript
speaker_lug string Orpheus speaker ID used for Luganda
speaker_eng string Orpheus speaker ID used for English
src_dur_s float Luganda clip duration after VAD trim (s)
tgt_dur_s float English clip duration after VAD trim (s)
dur_ratio float tgt / src duration ratio
src_speech_ratio float VAD speech fraction — Luganda
tgt_speech_ratio float VAD speech fraction — English
source_dataset string Originating HF dataset repo
source_subset string Originating subset / config name

Failure manifest

Rows that failed TTS generation (after 3 retries) or post-processing are logged to: https://huggingface.co/datasets/yigagilbert/synthetic-parallel-external/resolve/main/failure_manifest.jsonl