saudi-clean-vox

Saudi dialectal Arabic speech, cut and filtered for VoxCPM fine-tuning. 2132 single-speaker clips (3.47 hours) from 7 YouTube videos, mono 16 kHz 16-bit PCM WAV. Every clip carries a speaker label; there are 7 of them, one per source video, and they were not linked across videos -- treat two labels as two voices only within the video that produced them.

Built with youtube-to-dataset: diarization, VAD segmentation, music and overlap rejection, speaker-purity checks, SQUIM audio quality gates, and a two-transcript agreement check.

Layout

train.jsonl          # VoxCPM manifest, 2068 rows
val.jsonl            # VoxCPM manifest, 64 rows
wavs/*.wav           # 16000 Hz mono 16-bit PCM
wavs/metadata.jsonl  # HuggingFace audiofolder metadata
dataset_stats.json   # what was kept, what was dropped, and why

Clips run 1.76-12.38 s (mean 5.87 s), inside VoxCPM's usable range and well clear of the 1-second floor below which it becomes unstable.

Manifest rows look like this:

{"audio": "wavs/xxx_0000.wav", "text": "...", "duration": 6.1, "ref_audio": "wavs/xxx_0007.wav", "ref_duration": 5.2, "speaker": "xxx_SPK1", "video_id": "xxx"}

audio / text are what VoxCPM requires; duration and ref_duration let it skip decoding during length filtering; ref_audio is another clip of the same speaker, which trains the prompted (voice-cloning) path. Drop the ref_audio keys for plain text-to-speech fine-tuning.

Fine-tuning VoxCPM

Audio paths are relative to the dataset root, so either run training from that directory or rewrite them once:

huggingface-cli download Rabe3/saudi-clean-vox --repo-type dataset --local-dir saudi-clean-vox
cd saudi-clean-vox
python -c "
import json,pathlib
root=pathlib.Path.cwd()
for name in ('train.jsonl','val.jsonl'):
    rows=[json.loads(l) for l in open(name,encoding='utf-8')]
    for r in rows:
        for k in ('audio','ref_audio'):
            if k in r: r[k]=str(root/r[k])
    open(name,'w',encoding='utf-8').write(''.join(json.dumps(r,ensure_ascii=False)+chr(10) for r in rows))
"

Then point VoxCPM's config at them:

train_manifest: /path/to/saudi-clean-vox/train.jsonl
val_manifest:   /path/to/saudi-clean-vox/val.jsonl
sample_rate: 16000

python scripts/train_voxcpm_finetune.py --args.load conf/voxcpm_v2/voxcpm_finetune_lora.yaml

sample_rate must match the AudioVAE encoder input rate of the checkpoint you are fine-tuning (16000 Hz for VoxCPM 1.0 and VoxCPM 2; VoxCPM 1.5 wants 44.1 kHz, which this dataset cannot supply -- the source audio is 24 kHz).

Two things that bite: datasets>=4 needs torchcodec installed before it will decode an Audio column, and the longest sample here is about 550 audio frames plus text, so the stock max_batch_tokens: 8192 filters nothing out.

Loading it directly

from datasets import load_dataset
ds = load_dataset("Rabe3/saudi-clean-vox")                    # manifests, audio as paths
ds = load_dataset("audiofolder", data_dir="saudi-clean-vox/wavs")  # decoded audio + text

Transcripts

Every clip was transcribed twice: YouTube's auto-captions and Cohere Transcribe Arabic. Clips where the two disagreed by more than 35% CER were dropped. The text shipped here is the cohere transcript.

YouTube's Arabic auto-captions drop ta marbuta, hamza seats and almost all punctuation, which is fine for ASR and harmful for TTS -- a model trained on that spelling never learns to read ordinary written Arabic. Cohere's transcript preserves the orthography, and the agreement gate already established that the two say the same thing.

Transcripts are unvocalized (no harakat), as ordinary written Arabic is.

Quality gates

Every clip passed all of: single speaker, no overlapped speech, no music or applause (Demucs accompaniment energy + an AST AudioSet classifier), ECAPA speaker-purity against the speaker's own centroid, SQUIM MOS/STOI/PESQ, clipping and SNR bounds, and characters-per-second bounds calibrated for Arabic script.

Loudness is already normalized (about -16.4 LUFS across the set).

Licence and provenance

Audio is derived from YouTube videos and remains subject to YouTube's Terms of Service and the underlying copyright; video_id on each row keeps provenance attached. Parts of the pipeline that produced it (DiariZen, the MMS forced aligner) are CC BY-NC 4.0, so this dataset is released non-commercially.