Matoub-82M

Matoub-82M

A text-to-speech model for Kabyle (Taqbaylit, kab): an 82M-parameter StyleTTS2 fine-tune of Kokoro-82M trained on 21,953 restored Common Voice Kabyle clips from a fifties male speaker. It synthesises 24 kHz speech that reproduces the gemination, spirantisation, emphatics, and pharyngeals of Kabyle phonology in a native speaker voice — twice the sample rate of mms-tts-kab, the incumbent Kabyle TTS model, and under a licence that permits commercial use where that one does not.

The work that makes it Kabyle is the front end. Kokoro's token table maps 114 symbols and carries none of ˤ ʕ ħ — pharyngealisation, and the letters ɛ and ḥ — while its G2P is built with unk='' and drops a phoneme it cannot represent rather than raising, which would have deleted three consonants from every training target behind a healthy loss curve. The 42-symbol Kabyle inventory was diffed against that table, the three missing symbols assigned to unused embedding rows and trained, the affricate tie bar folded onto ʧ and ʤ, and the front end made to validate against the vocabulary and raise. That is why the emphatics and pharyngeals survive to the decoder.

One voice is published here — kab_male. The female Stage 2 and the Cycle-CER are the next two, and this repository carries no claim about either until they land.

Named after Lounès Matoub (1956-1998), Kabyle singer, poet, and tireless voice of Taqbaylit, who gave his life to its language and culture.

Results

The baseline for Kabyle TTS is mms-tts-kab (Meta's MMS). Cycle-CER measures the acoustic distortion introduced by synthesis: synthesise, transcribe with agbalu/Fadhma-300M, measure character error rate against the original text.

That baseline row is itself a result of this project. mms-tts-kab had been published without a Kabyle score by anyone; the +3.56 delta was measured here on 2026-08-14 over 1,000 held-out non-biblical prompts, against a real-audio control on the same text. Kabyle TTS now has a number to be judged against, and this model will be judged against it too — the run has not been made, so the row stays empty rather than estimated.

Three things worth reading carefully.

The baseline is read scripture at 16 kHz. mms-tts-kab is MMS's per-language VITS checkpoint, trained on New Testament recordings — which typically carry one speaker per language — and it emits 16 kHz under cc-by-nc-4.0. Matoub-82M was fine-tuned on Common Voice read speech from a native Kabyle male speaker and emits 24 kHz under Apache-2.0. No listening test has been run between the two, and none is claimed here.

The training audio has a hard frequency ceiling. The kab_male clips are band-limited at approximately 7.9 kHz -- not 11.5 kHz or 24 kHz -- because the recording conditions for Common Voice Kabyle combined with phone microphones, lossy encoding, and upload artefacts cut the spectral content. The model cannot synthesise what was not in its training data; any evaluation above 7.9 kHz measures silence. This is a property of the Kabyle speech record rather than of this checkpoint — the incumbent synthesises at 16 kHz, so both systems are band-limited, and closing it needs recordings that do not currently exist.

Diffusion was never trained, so it is not in the release. lambda_diff: 0.0 for every epoch, which left the style sampler exactly as Kokoro initialised it — blending its output into the decoder injects noise from a module that learned nothing. The export drops it, along with the two style encoders, and carries the speaker style as a 256-dim vector instead. There is no alpha or beta to set and no way to set one wrong.

The vocoder is stochastic. The harmonic-plus-noise source draws its noise floor and initial phase from the global RNG, so two calls on the same text return different samples. Seed torch if you need one waveform twice.

Intended use

Producing spoken Kabyle from text for:

• Accessibility: screen readers and audio production for Kabyle-language content.
• Language learning: audio for learners studying Taqbaylit.
• NLP pipeline completion: the terminal stage of a full Kabyle text pipeline, downstream of agbalu/Juba-27M (Tifinagh to Latin), agbalu/Belaid-31M (punctuation and casing), and agbalu/Boulifa-48M (orthography standardisation).

Not suitable for: any use requiring speaker consent or biometric match to the source speakers; cloning the voice of any person who has not consented; any decision about a person; any language other than Kabyle. No safety evaluation of any kind has been performed.

What comes next

This checkpoint establishes the parts that carry forward: a Kabyle phoneme inventory the base model can represent, a G2P front end that raises instead of dropping, a two-stage recipe that trains to a monotone validation curve on 17.9 hours, and an end-to-end path from Kabyle text to 24 kHz audio.

Next, in order: the kab_female Stage 2, the Cycle-CER against the mms-tts-kab baseline above, then UTMOSv2 and speaker similarity.

This repository stays permanently published at this URL. Voices that follow are released under their own names as standalone repositories rather than replacing it.

Usage

pip install "transformers>=5" torch soundfile

import soundfile as sf
from transformers import AutoModelForTextToWaveform, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("agbalu/Matoub-82M", trust_remote_code=True)
model = AutoModelForTextToWaveform.from_pretrained("agbalu/Matoub-82M", trust_remote_code=True)
model.eval()

inputs = tokenizer("Azul fell-awen, amek i telliḍ taṣebḥit-a?", return_tensors="pt")
audio = model(**inputs).waveform[0]

sf.write("output.wav", audio.numpy(), model.config.sampling_rate)

trust_remote_code=True is required: StyleTTS2 is not a transformers architecture, so the repository ships the modelling code beside the weights and nothing else is needed to run it — no StyleTTS2 checkout, no reference clip, no separate G2P package.

forward takes speed (a duration multiplier, default 1.0) and voice (a 256-dim style vector, defaulting to the one in the weights). It returns waveform, waveform_lengths and durations — frames per input phoneme, which is the alignment. A padded batch is synthesised item by item and right-padded; read waveform_lengths before trimming.

Text in, phonemes out. The tokenizer is the front end: it runs the Kabyle G2P, folds the tie-bar affricates (t͡ʃ → ʧ, d͡ʒ → ʤ) onto the symbols in the token table, and raises on a character it has no rule for rather than dropping it. Inspect what it produced with tokenizer.phonemize(text). Punctuation is discarded and the clitic hyphen is a word boundary, because that is what the training transcripts carried — the model has never been supervised on a comma. PL-BERT positions 510 tokens, so synthesise one sentence at a time.

The synthesis pipeline:

1. G2P — Kabyle Latin orthography to IPA, with gemination, spirantisation, a-backing and nasal assimilation.
2. Duration and pitch prediction — bert (PL-BERT, 12 layers), bert_encoder and predictor predict per-phoneme durations and the F0 and energy contours from the token sequence and the prosodic half of the style vector.
3. Waveform decoding — text_encoder's features are expanded to frames by the predicted durations, and the iSTFTNet decoder renders 24 kHz mono audio conditioned on the acoustic half.

Architecture

Matoub-82M is a StyleTTS2 model initialised from Kokoro-82M weights and fine-tuned in two stages:

The training checkpoint also held the two style encoders, the style diffusion sampler, the text aligner, the JDC pitch extractor, the MPD and MSD discriminators, the WavLM discriminator and the optimizer state — 1.93 GB of instruments that produce nothing at inference. None of them is in the release. The style encoders' one output, the 256-dim speaker vector, is; it was extracted from a restored kab_male corpus clip on the filterbank the recipe trains on, 80 mel bins over n_fft 2048 and hop 300 with f_max 8 kHz.

Stage 1 (multi-speaker) trains text_encoder, style_encoder, decoder, mpd, msd on both voices merged under global speaker IDs. It builds the acoustic backbone from the Kokoro base.

Stage 2 (per-voice) freezes the Stage 1 acoustic modules and fine-tunes bert, bert_encoder, predictor, predictor_encoder -- the language-model and duration stack -- on one voice at a time. It is where Kabyle prosody and phoneme timing are learned.

Training data

Corpus: Common Voice Kabyle, restored arm. Clips were amplitude-normalised, silence-trimmed, and filtered: flat-topped (clipped) samples and zero-energy clips were removed entirely.

The audio quality of kab_male is the binding constraint on this checkpoint. Inspection of the spectrograms shows the signal cut off at approximately 7.9 kHz, consistent with recording on a smartphone through a codec that discards high frequencies before upload. Of the 3.73 s mean clip, 23% is silence. The clips supervise no pause structure and no multi-sentence prosody: the longest clip is 10.5 s and fewer than 1.2% reach 8 s.

This is not a flaw in the data preparation -- it is a measurement of what Common Voice Kabyle recordings contain, and every claim this model makes about audio quality should be read against it.

Training recipe

Stage 1 -- multi-speaker acoustic pretraining:

Stage 1 validation curve (monotone, decelerating -- the last two epochs bought 0.002 each):

Stage 2 -- per-voice language-model fine-tuning (kab_male):

make modal-matoub TASK=prepare ARM=restored
make modal-matoub TASK=stage2 ARM=restored VOICE=kab_male EPOCHS=5

Stage 2 (kab_female) is not yet published. The female voice requires a separate Stage 2 run; the checkpoint published here covers only the male voice.

Limitations

The kab_male recording quality defines the quality ceiling. Phone microphone recordings at ~7.9 kHz effective bandwidth, with 23% silence per clip and a maximum clip length of 10.5 s, are the training distribution. The model cannot exceed what it was shown. The frequency ceiling is the most consequential limitation: 24 kHz output with nothing above 7.9 kHz is broadband silence from 7.9 kHz upward, and it will be audible on any speaker or headphone that reproduces it.

Single published voice. The kab_female Stage 2 has not been trained to a publishable checkpoint. The card will be updated when it completes.

Short-clip corpus. Mean clip length is 3.73 s (male) and 3.11 s (female), of which 23-43% is silence. The model has not been supervised on multi-sentence prosody, pause structure, or paragraph-level intonation. Long sentences are synthesised phoneme-by-phoneme; paragraph rhythm is not modelled.

One language. Trained and evaluated on Kabyle. Tarifit, Tashelhit, Central Atlas Tamazight and Shawiya have related but distinct phonologies; none was tested and none should be assumed.

No safety evaluation of any kind has been performed.

What the numbers cover, and what they do not

The mms-tts-kab row in Results is measured. Everything below is open, listed so you can plan around it:

• This model's own Cycle-CER. The instrument and the baseline both exist; the run has not been made.
• MOS and UTMOSv2. No perceptual evaluation has been performed.
• Speaker similarity. No embedding comparison against the source voice has been performed.
• Long-form synthesis. Degradation on multi-sentence or paragraph-length input is unquantified, and the training clips give a strong prior that it exists — see Limitations.
• The female voice. kab_female Stage 2 has not been trained.

Files

Nothing else is needed to run it: no StyleTTS2 checkout, no reference clip, no separate G2P package. Training cannot be resumed from these files — the optimizer, the discriminators and the style diffusion sampler stay in the 1.93 GB source checkpoint on the training volume, because none of them produces anything at inference.

Reproduction

make modal-matoub TASK=pull                       # the training checkpoint
make modal-matoub-reference                       # list the voice's restored clips
make modal-matoub-reference CLIP=<the first>.wav  # the clip the speaker style is taken from
make push REPO=matoub                             # export, stage, verify, upload

The export bakes the speaker style into the weights, and that style is a function of one restored clip — the alphabetically first in the voice's directory. Its name is a Common Voice clip id and is not derivable, which is why the listing step exists.

Full training reproduction:

make modal-matoub TASK=prepare ARM=restored
make modal-matoub TASK=stage1 ARM=restored EPOCHS=6
make modal-matoub TASK=stage2 ARM=restored VOICE=kab_male EPOCHS=6

The name

Lounès Matoub (1956-1998) was the most celebrated Kabyle singer of the 20th century and one of the fiercest advocates for the survival of Taqbaylit. He recorded 36 albums in Kabyle at a time when the Algerian state was suppressing Berber language and culture, making the language audible to an entire generation. He was assassinated on 25 June 1998, ten days before the Arabisation law he had spent years opposing took effect.

His voice is inseparable from the survival of Kabyle as a spoken language in collective memory. Naming a Kabyle TTS model after him is not metaphor; it is acknowledgment that what this model does -- make the language heard -- is what he spent his life doing.

The naming is homage; it implies no endorsement by anyone.

Citation

@software{agbalu_matoub_2026,
  title  = {Matoub-82M: neural speech synthesis for Kabyle},
  author = {AƔBALU},
  year   = {2026},
  url    = {https://huggingface.co/agbalu/Matoub-82M},
  note   = {StyleTTS2 fine-tune of Kokoro-82M on 21,953 restored Common Voice Kabyle
            clips; 24 kHz, Apache-2.0, one published voice}
}

Licence

Apache-2.0 on the weights and the code. The training data derives from Common Voice Kabyle (CC0); the Kokoro base weights are published under Apache-2.0. A permissive grant on weights makes no claim about the voice recordings they were trained on.

Part of AƔBALU, a Kabyle corpus and model collection.