alconost/alconost-multilingual-speech-gold

Multilingual Speech & Translation Dataset — EN↔JA/AR-EG/PL/RU (10 phrases, dual-take)

Description 10 English source phrases with expert human translations into Japanese, Egyptian Arabic (ar-EG), and Polish. Each target phrase is recorded by native speakers (two takes each). Audio files are WAV 48 kHz mono, 16‑bit PCM format. Translations are produced and QA'd by professional linguists; recordings follow consistent orthography/style (AR-EG: Egyptian dialect; JA/PL: standard). All speakers are native Egyptian, Japanese, or Polish speakers from their respective regions. Each row carries an explicit consent flag.

Contents

• 10 source phrases (English)
• 1 unified metadata CSV with 330 rows (one row per audio file)
• 330 audio files
• Speaker demographics (gender, age) and consent info

Per-language breakdown (this release)

• ar-EG: 100 audio files (5 speakers × 10 phrases × 2 takes)
• pl: 100 audio files (5 speakers × 10 phrases × 2 takes)
• ja: 80 audio files (4 speakers × 10 phrases × 2 takes)
• ru: 50 audio files
  • 2 speakers × 10 phrases × 2 takes (40 files)
  • 1 speaker × 10 phrases × 1 take (10 files)
  • Note: spk_ru_002 take2 is labeled background_noise_level=Loud; all other audio is Quiet.

Use Cases Multilingual TTS/ASR fine‑tuning, cross‑lingual speech translation, multimodal alignment.

License & Consent Each translation row includes consent (TRUE in this release). Optional forms: /consent/{speaker_id}/.

Audio Specs All audio files: WAV 48 kHz mono, 16-bit PCM format, normalized to −23 LUFS ±2 LU.

All 280 audio files have been normalized to meet the target loudness specification (−23 LUFS ±2 LU). Sample rate and channel configuration (48 kHz mono) are consistent across all files.

Quality & Provenance Professional translation with bilingual QA; native‑speaker recordings (studio conditions); two takes per item; terminology/orthography consistency; verified consent.

Data layout

/alconost-multilingual-speech-en-ja-ar-pl-v1/
├── README.md
├── LICENSE.txt
├── metadata.csv
└── audio/
    ├── seg_0001_ar_spk_ar_eg_001_take1.wav
    ├── seg_0001_ar_spk_ar_eg_001_take2.wav
    ├── seg_0001_ja_spk_ja_001_take1.wav
    └── ... (280 audio files total)

Audio files are stored in a flat structure. File naming format: seg_{####}_{lang}_spk_{speaker_id}_take{1|2}.wav

CSV schema

• metadata.csv: file_name, id, source_id, source_text, target_lang, target_text, domain, speaker_id, gender, age, consent, background_noise_level
  • file_name: Audio file path (relative to dataset root) - required by Hugging Face Dataset Viewer
  • id: Unique identifier for each translation recording (e.g., seg_0001_ar_spk_ar_eg_001)
  • source_id: Reference to source phrase (e.g., seg_0001)
  • source_text: English source text
  • target_lang: Target language code (ar-EG for Egyptian Arabic, ja for Japanese, or pl for Polish)
  • target_text: Translated text in target language
  • domain: Domain category (Daily life, Technology, Weather, Health, Travel, Education, Food, Business, Entertainment, Emergency)
  • speaker_id: Anonymous speaker identifier (e.g., spk_ar_eg_001, spk_ja_001, spk_pl_001)
  • gender: Speaker gender (male or female)
  • age: Speaker age (numeric)
  • consent: Consent flag (TRUE for all entries)
  • background_noise_level: Background noise level (Quiet or Loud)

Citation

Alconost, 2025. “Multilingual Speech & Translation Dataset (EN↔JA/AR-EG/PL).”

Alconost Linguistic Data Labeling Services

Alconost provides high-quality linguistic data labeling for AI teams worldwide — including parallel corpora, parallel texts, error annotation, multilingual glossaries, LQA/LQT, and finely curated audio and video datasets. Leveraging 20+ years of linguistic expertise, native-speaking specialists, and AI-enhanced workflows, we build clean, scalable datasets for ASR/TTS, LLMs, chatbots, and multilingual NLP. From annotation to QA, we help you train models that understand nuance, context, and dialect variation.

Contact: ai-data@alconost.com to order or purchase datasets.