GABAR

Open source

Open voices for Ethiopian languages

We fine-tune existing speech models for Ethiopian languages using licensed and publicly available datasets. We publish the adapters, evaluations, text front-ends, and training-data audits under CC-BY-SA-4.0.

Hugging FaceGitHubTry it in the browser

Before and after

The same sentence, the same voice, with and without the adapter.

Each comparison uses the same text, reference recording, and generation settings. The clips are single, unselected outputs.

Chatterbox Amharic

An adapter that teaches Chatterbox Multilingual v3 to speak Amharic, with voice cloning from about ten seconds of reference audio. The base model has no tokens for the Ge'ez script at all, so the "before" clips are the model guessing.

193 hours from three sources, 244 new tokens for the Ge'ez script, 100 held-out clips from speakers the model never heard.

Model and weightsSource

Ordinary prose

ኢትዮጵያ የረጅም ዘመን ታሪክና ኩሩ ባህል ያላት ጥንታዊት ሀገር ናት።

0:00 / 0:00

Numbers and a date abbreviation

በ2016 ዓ.ም. የአገሪቱ የሕዝብ ብዛት ከ120 ሚሊዮን በላይ እንደደረሰ ተገምቷል።

0:00 / 0:00

A title abbreviation

ዶ/ር አበበ በአዲስ አበባ ዩኒቨርሲቲ የቋንቋ ጥናት ያስተምራሉ።

0:00 / 0:00

A question

አሁን የሚሰሙት ድምፅ ከፅሑፍ በቀጥታ የተፈጠረ መሆኑን ሲያውቁ የድምፅ ቃናው ምን ያህል ተፈጥሯዊ ነው?

0:00 / 0:00

Numbers, percent, punctuation

ቁጥሮች ለምሳሌ 1500 ወይም 75% እንዴት ይነበባሉ? የስርዓተ ነጥብ አጠቃቀምስ እንዴት ይስተናገዳል?

0:00 / 0:00
MetricStock Chatterbox v3Gabar adapter
Character error rate (lower is better)0.9320.095
Speaker similarity, ECAPA cosine (higher is better)0.6100.860
UTMOS naturalness predictor2.3592.711
Generation failures0 %1 %

Chatterbox Afaan Oromo

The same approach for Afaan Oromo. Qubee is Latin script, so the base model can read the letters — what it lacks is the language, and it reads them with a foreign accent.

210 hours from two sources, 100 held-out clips from speakers the model never heard.

Model and weightsSource

Ordinary prose

Itoophiyaan biyya seenaa dheeraa fi aadaa boonsaa qabdu dha.

0:00 / 0:00

Numbers and a year

Bara 2018 keessa baay'inni ummata biyyattii miiliyoona 120 ol akka ga'e tilmaamameera.

0:00 / 0:00

Names and ejectives

Doktar Caaltuun Yunivarsiitii Finfinnee keessatti qorannoo afaanii barsiisti.

0:00 / 0:00

A question

Sagaleen amma dhageessan kun barreeffama irraa kallattiin kan uumame ta'uu isaa yoo beektan, sagaleen kun hammam uumamaa fakkaata?

0:00 / 0:00

Numbers, percent, punctuation

Lakkoofsonni, fakkeenyaaf 1500 yookiin 75%, akkamitti dubbifamu? Mallattooleen tuqaas akkamitti keessummeeffamu?

0:00 / 0:00
MetricStock Chatterbox v3Gabar adapter
Character error rate (lower is better)0.4580.193
Speaker similarity, ECAPA cosine (higher is better)0.7140.809
UTMOS naturalness predictor2.3572.531
Generation failures0 %0 %

CosyVoice2 Amharic

A fine-tune of the text-to-token and token-to-mel stages of CosyVoice2-0.5B for Amharic, keeping the base model’s voice cloning from about ten seconds of reference audio. The stock checkpoint did not produce intelligible Amharic in our tests.

194 hours from three sources, 100 clips evaluated from a 104-clip held-out split of speakers the model never heard. The fine-tune inserts an audible hesitation before some words — its main quality limitation.

Model and weightsSource

Ordinary prose

ኢትዮጵያ የረጅም ዘመን ታሪክና ኩሩ ባህል ያላት ጥንታዊት ሀገር ናት።

0:00 / 0:00

Numbers and a date abbreviation

በ2016 ዓ.ም. የአገሪቱ የሕዝብ ብዛት ከ120 ሚሊዮን በላይ እንደደረሰ ተገምቷል።

0:00 / 0:00

A title abbreviation

ዶ/ር አበበ በአዲስ አበባ ዩኒቨርሲቲ የቋንቋ ጥናት ያስተምራሉ።

0:00 / 0:00

A question

አሁን የሚሰሙት ድምፅ ከፅሑፍ በቀጥታ የተፈጠረ መሆኑን ሲያውቁ የድምፅ ቃናው ምን ያህል ተፈጥሯዊ ነው?

0:00 / 0:00

Numbers, percent, punctuation

ቁጥሮች ለምሳሌ 1500 ወይም 75% እንዴት ይነበባሉ? የስርዓተ ነጥብ አጠቃቀምስ እንዴት ይስተናገዳል?

0:00 / 0:00
MetricStock Chatterbox v3Gabar adapter
Character error rate (lower is better)0.9390.094
Speaker similarity, ECAPA cosine (higher is better)0.3970.758
UTMOS naturalness predictor2.1772.787
Generation failures0 %0 %

All eight pairs and the reference recording are on each model page.

Measured

Held-out evaluation, speakers the model never heard.

Same clips, same per-clip reference audio, same code path for both models. Character error rate is judged by Meta's stock omniASR-CTC-3B, which we did not train. UTMOS is an English-trained naturalness predictor and these outputs are not English, so treat it as a rough signal.

Per-clip results, training-data audits, held-out evaluation details, and watermark checks are included with each release.

What went in

Licensed data, an open text front-end, and a delta rather than a copy.

Licensed data only

Our own studio recordings, WaxalNLP (Digital Umuganda / Google, CC-BY-SA-4.0), Common Voice (CC0), and a published research corpus for Afaan Oromo. Every clip carries its corpus of origin; the assembled training directory is audited before training and the audit ships with the model.

An adapter, not a copy

We publish our delta — LoRA on the text-to-speech transformer plus the new embeddings — and you apply it to the base model, which you get from its authors. Everything else stays frozen.

An open text front-end

Numbers, dates, abbreviations, spellings and punctuation are normalized by a single dependency-free file per language. It ships with the model, it built the training labels, and it works on its own.

Watermarked, share-alike

Every output carries Resemble's PerTh watermark; we left it intact and verified it. Because WaxalNLP is share-alike, our releases are CC-BY-SA-4.0 too, with credit to the data's authors.

What these are not

These are adapters for Chatterbox Multilingual v3, not voice models built from scratch. Voice cloning comes from the base model; our work fine-tunes it for Amharic and Afaan Oromo. The results are not yet fully natural, and each model card documents the evaluation and known limitations.

Read the model cardsTalk to us