AfriSpeech/youversion-african-speech-prepped

YouVersion African Speech — prepped for MOSS-TTS-Nano finetuning

Preprocessed training data for finetuning MOSS-TTS-Nano: every audio clip from AfriSpeech/youversion-african-speech has been encoded into discrete audio codes with MOSS-Audio-Tokenizer-Nano, so you can train directly with the MOSS-TTS-Nano finetuning pipeline without downloading or re-encoding the ~63 GB of source audio.

What's in it

• 93,414 records across all 50 language subsets of the source dataset (Bible audio segments from YouVersion).
• Plain text-to-speech pairs — no reference/voice-cloning audio: this data trains the model's unconditional (single-voice) TTS path. Note the source audio has one narrator per language, so a model trained on the full set hears ~50 different voices.
• Split into 6 JSONL shards under youversion-african-speech-all-seed0/: train_with_codes.rank0000K-of-00006.jsonl (15,569 records each, interleaved row-robin from a seed-0 shuffle of the full export).

Record format

One JSON object per line:

{
  "text": "Mʼagya te ase anaa?",
  "language": "Akan",
  "audio": "/tmp/ft_data/wav/Akan_aka_000123.wav",
  "audio_codes": [[512, 87, ...], ...]
}

Field	Description
text	Transcript of the clip (the TTS target text)
language	Language name from the source dataset, e.g. Akan, Amharic, Yoruba — injected into the training prompt's Language: field
audio	Original wav path from the preprocessing container — stale, ignore it; audio_codes is what training consumes
audio_codes	(T, 16) integer codes from MOSS-Audio-Tokenizer-Nano (16 codebooks, ~12.5 frames/sec)

How it was made

From the source dataset's train splits, with these filters:

• duration within 0.5–30 s (5,498 clips dropped)
• text length ≤ 400 characters (8 clips dropped)
• audio resampled by the codec loader to the tokenizer's sample rate; encoded with finetuning/prepare_data.py --skip-reference-audio-codes from the MOSS-TTS-Nano repo

Heads-up: a small number of records (a handful, mostly Amharic/Tigrinya — Ge'ez script tokenizes at >2 text tokens per character) pack to more than 1024 tokens. If you train with --max-length 1024, filter those out first or dataset.py will raise on prompt overflow.

How to train with it

# 1. Get the finetuning code
git clone https://github.com/OpenMOSS/MOSS-TTS-Nano
cd MOSS-TTS-Nano && pip install -r requirements.txt

# 2. Download the shards
hf download AfriSpeech/youversion-african-speech-prepped --repo-type dataset --local-dir prepped

# 3. Train (no prepare_data.py needed — codes are precomputed)
accelerate launch finetuning/sft.py \
  --model-path OpenMOSS-Team/MOSS-TTS-Nano-100M \
  --codec-path OpenMOSS-Team/MOSS-Audio-Tokenizer-Nano \
  --train-jsonl 'prepped/youversion-african-speech-all-seed0/train_with_codes.rank*.jsonl' \
  --output-dir output/moss_tts_nano_youversion \
  --per-device-batch-size 4 --gradient-accumulation-steps 4 \
  --learning-rate 1e-5 --num-epochs 3 --mixed-precision bf16 \
  --max-length 1024 --channelwise-loss-weight 1,32

To train on a subset of languages, filter the JSONL by the language field first.

A finetuned checkpoint trained on exactly this data is published at AfriSpeech/moss-tts-nano-youversion-sft.

Attribution & license

Derived from AfriSpeech/youversion-african-speech (Bible audio segments from YouVersion). Use of this derivative is subject to the same terms as the source dataset — please consult the source dataset card before using it beyond research.