KFUPM-JRCAI/arabic_speech

Arabic Speech Dataset

A curated Arabic speech dataset combining three sources, transcribed via Whisper-large-v3 (primary) and omniASR-7B (fallback). Built for Arabic TTS and ASR training.

Quick Start

from datasets import load_dataset

ds = load_dataset("KFUPM-JRCAI/arabic_speech", split="train")
print(ds[0]["transcription"])         # ASR text
print(ds[0]["audio"]["array"])        # audio waveform (float32, 16kHz)
print(ds[0]["timestamps"])            # word-level timestamps (nullable)
print(ds[0]["model"])                 # "whisper-large-v3" or "omniASR-7B"
print(ds[0]["origin"])                # source dataset name
print(ds[0]["speakerandsession"])     # "session_id::speaker"

Schema

Column	Type	Description
audio	Audio(16000)	Decoded mono waveform at 16kHz
transcription	string	Raw ASR output (not Gemini-normalized)
model	string	"whisper-large-v3" or "omniASR-7B"
timestamps	List[{word, start, end}]	Word-level timestamps (nullable)
speakerandsession	string	{session_id}::{speaker}
origin	string	Source HF dataset name

Data Sources

Source	Origin	Rows	Description
ArabicVoicesClean_v5	KFUPM-JRCAI/ArabicVoicesClean_v5	2,961	Crowd-sourced Arabic speech recordings
Miro	TigreGotico/tts-train-synthetic-miro_ar-diacritics	237	TTS-synthetic Arabic with diacritics
DII	TigreGotico/tts-train-synthetic-dii_ar-diacritics	4,350	TTS-synthetic Arabic with diacritics

Total: 7,548 rows (6,352 whisper / 1,196 omniASR).

How It Was Built

The pipeline starts from the original datasets (audio + text pairs), then applies automatic speech recognition (ASR) to filter and enrich the data:

1. ASR Transcription & Filtering

For each audio-text pair from the original datasets, we run Whisper-large-v3 (and omniASR-7B as a fallback) to produce an ASR transcript. The ASR transcript is then compared against the original text:

• Rows where the ASR output matches the original text well are marked keep=true in filtered_records.jsonl -- these are the high-quality pairs retained in this dataset.
• Rows where the ASR deviates significantly are discarded. The original text may contain errors (e.g., mismatched audio), and the ASR serves as a quality gate.

2. Word-Level Timestamps

For retained rows, whisper also produces word-level timestamps (word, start, end) stored in asr_words_cache.jsonl. These enable alignment tasks and fine-grained analysis.

3. Union (Whisper-preferred)

When both whisper and omniASR transcriptions exist for a row, the whisper version is used. omniASR is only used when whisper coverage is missing (1,196 rows).

4. Clip ID Mapping

HuggingFace Dataset loads wav files in alphabetical order (1.wav, 10.wav, 100.wav, ...), so a naive ds_idx + 1 mapping maps audio to the wrong transcription. To fix this, each row's original text is matched against metadata.csv to find the real clip_id.

5. Audio Loading

• ArabicVoicesClean_v5: Audio decoded from parquet shards (embedded binary WAV) via soundfile.read at 16kHz.
• Miro/DII: Audio loaded from complete wav directories at /tmp/tts-train-synthetic-*-hf/wav/.

Known Limitations

Missing Word Timestamps

The word-level timestamp generation (asr_words_cache.jsonl) was only run for a subset of indices:

• Miro: indices 1-383 (237/570 keep rows have timestamps)
• DII: indices 1-7,943 (4,333/5,559 keep rows have timestamps)
• ArabicVoicesClean_v5: 100% covered

How to fix: Re-run the word-segmentation step on the remaining indices. The asr_text_cache.jsonl has full coverage -- only the timestamp extraction was interrupted.

Incomplete Miro Coverage

The miro ASR was only run on the first 968 clip_ids (indices 0-967) out of ~9,994 total. To expand coverage, re-run ASR inference on the remaining rows and include them in the keep set.

License

Refer to the licenses of the individual source datasets:

• KFUPM-JRCAI/ArabicVoicesClean_v5
• TigreGotico/tts-train-synthetic-miro_ar-diacritics
• TigreGotico/tts-train-synthetic-dii_ar-diacritics