MoulSot-Tokens-v1

Discrete speech tokens for Moroccan Darija. 101 hours of transcribed speech, encoded to a single-codebook neural audio codec and paired with text, ready to train a text-to-speech model that predicts tokens directly.

	Size
Source audio (16 kHz parquet)	~12 GB
This dataset (tokens, text)	153 MB

Same speech, ~78× smaller. At the codec level that is 256 kbps of PCM reduced to 0.8 kbps — a 320× reduction in bits — since in this case, one second of audio becomes 50 integers instead of 16,000 samples.

That is the point of this release. Darija TTS work previously required downloading tens of gigabytes of audio and running a GPU encode pass before any modelling could start. This dataset removes that step.

The goal of this dataset is to accelerate R&D in using text to speech models on Moroccan Darija (Tamazight on the radar as well). This release provides an ultra small dataset, with over 100 hours of speech data tokenized.

---

At a glance

Clips	79,641
Duration	101.6 hours (mean 4.6 s, median 3.5 s)
Tokens	18.3 million @ 50 tokens/second
Codec	neuphonic/neucodec — FSQ, 1 codebook, 65,536 codes
Source	atlasia/MoulSot-Full, config 100-gt-2.5
Transcripts	Generated with Gemini 2.5 Pro (see source dataset card)
License	Apache-2.0

Single codebook. Most neural codecs use residual vector quantisation with 8–12 codebooks, so a second of audio becomes 600 tokens. NeuCodec uses one FSQ codebook at 50 Hz, which means a sequence model sees 50 tokens per second.

---

Usage

from datasets import load_dataset

ds = load_dataset("Tilas/MoulSot-Tokens-v1", split="train")
row = ds[0]

row["text"]     # الترانسكريبت بالدارجة
row["tokens"]   # [1243, 8421, 9901, ...]  ints in [0, 65536)
len(row["tokens"]) / 50   # duration in seconds

Listening to a sample

import torch
from neucodec import NeuCodec

model = NeuCodec.from_pretrained("neuphonic/neucodec").eval()
codes = torch.tensor(row["tokens"]).reshape(1, 1, -1)
audio = model.decode_code(codes)      # 24 kHz — the decoder has no 16 kHz path

---

Schema

Field	Type	Notes
audio_id	string	Matches the source dataset
text	string	Darija transcript
channel	string	Source YouTube channel
duration	float	Seconds
n_samples	int	Raw 16 kHz sample count; n_samples // 320 == len(tokens)
tokens	list[int]	FSQ codes in [0, 65536), 50 per second
encoder	string	neucodec-v1
sampling_rate	int	16000 (encoder input rate)

---

Reproducing the tokens

Token IDs depend on exactly what the encoder receives. Two details are easy to get wrong and both silently change the output — we measured each:

Do not pre-pad the audio. NeuCodec pads internally. Appending your own padding first shifts the semantic encoder's per-utterance feature normalisation: 10 ms of added silence changed 85.3% of tokens.

Do not batch clips of different lengths. encode_code applies no attention mask, so padding a short clip up to a longer batch-mate perturbed 86% of its frames, starting at frame 0. Batch only clips with identical sample counts, or use batch size 1.

model = NeuCodec.from_pretrained("neuphonic/neucodec").eval()
# wav: 1-D float32 @ 16 kHz, raw length, mono, NO padding
codes = model.encode_code(wav.reshape(1, 1, -1))

Frame count is exactly n_samples // 320. This held for 79,857 / 79,857 clips in the full encode — the n_samples column lets you verify any row without touching audio.

---

Quality audit

3,090 clips were decoded and compared against their originals. Full methodology and per-clip data are in the project repository.

Speaker identity is retained

ECAPA cosine similarity between original and reconstruction averages 0.709 on clips ≥2 s.

ECAPA's scale. Speaker verification with ECAPA-TDNN separates same-speaker from different-speaker around 0.25–0.35 cosine. Different speakers land near 0.1–0.3. Two genuine recordings of the same speaker, from different sessions, typically land around 0.5–0.7 — microphone, room and session variation move the embedding that much on their own. A score of 0.709 sits in same-speaker territory

English gives the same number. Running the identical pipeline on LibriSpeech returns 0.723. If 0.709 signalled damage to Darija, English would not land in the same place. Both languages sit together, which means the number reflects what 0.8 kbps can carry, not how the codec handles Darija.

Intelligibility: seen data vs unseen data

Metric	Original	Reconstruction
WER	9.57	37.83
CER	3.71	18.18

A 28-point gap looks like severe damage. It is not — it is a seen/unseen effect.

The ASR used (atlasia/moulsot.v0.3) was fine-tuned on this same audio. Every original in this audit is a clip the model already trained on, so 9.57 WER is a memorisation score The model's own published performance on data it has not seen is 38.99 WER.

A reconstruction is acoustically different from its original, the same speech, resynthesised, so it no longer matches what the model memorised. From the ASR's point of view it is unseen data. It scores 37.83: the published baseline.

Basically The codec moved the audio from seen to unseen, and on unseen audio the model performs exactly as documented.

Native Darija speakers confirmed reconstructions are clearly intelligible. A clean measurement on held-out audio is in progress.

Signal quality

Mel-L1 averages 0.513. SI-SDR averages −8.94 dB, which sounds alarming and is not: NeuCodec decodes through a Vocos head that reconstructs perceptually and does not preserve phase, so waveform-domain metrics report large errors on audio that is perceptually intact.

---

Limitations

channel is not a speaker label. One channel can host many speakers and one speaker can appear across channels. Do not use it for multi-speaker conditioning without diarisation.

Transcripts are machine-generated Expect transcription noise.

---

Provenance

Audio and transcripts come from atlasia/MoulSot-Full, released under Apache-2.0.

Tokens were produced with the full neuphonic/neucodec encoder, using the entire encode path (Wav2Vec2-BERT semantic encoder, BigCodec acoustic encoder, FSQ quantiser) frozen. No weights were modified, so these tokens reproduce against the upstream checkpoint.

Citation

@misc{moulcodec2026,
  title  = {MoulSot-Tokens-v1: Discrete Speech Tokens for Moroccan Darija},
  author = {Aymen Echchalim},
  year   = {2026},
  url    = {https://huggingface.co/datasets/Tilas/MoulSot-Tokens-v1}
}

Built as open infrastructure for the Moroccan AI community. Issues and contributions welcome.