chatterbox-amharic

Chatterbox Amharic

A LoRA adapter and an extended Fidel tokenizer that teach Chatterbox Multilingual v3 (Resemble AI, MIT) to speak Amharic, with voice cloning from about ten seconds of reference audio. Trained only on speech we own or that is licensed for it.

Stock Chatterbox cannot read Amharic at all: its tokenizer maps every Ge'ez character to [UNK]. So the "before" clips below aren't a weaker version of the same thing; they're the model guessing at unknown tokens. We add 244 tokens for the script and teach the model what they sound like.

The repo also has amharic_text.py, the text normalizer the model was trained through. No dependencies, works on its own (below).

Hear it

Each pair uses the same sentence, reference audio, settings and seed. No language tag on either: the adapter was trained without one, and the stock tokenizer has no Ge'ez characters.

Reading this on GitHub? The players only render on Hugging Face. Click a clip in demo/ to play it, or watch demo/before_after.mp4 (all eight pairs, 1:47). The weights (new_lang_adapter/, 194 MB) are only on Hugging Face; everything else is mirrored here.

Texts: demo/sentences.txt. Reference voice: demo/reference.wav, one of us (<audio controls preload="none" src="https://huggingface.co/gabar-tech/chatterbox-amharic/resolve/main/demo/reference.wav"></audio>). Both models got the text after amharic_text.normalize, temperature=0.6, cfg_weight=0.5, seed 1234. One take per sentence per model, no picking. Known weaknesses are under Limitations.

What this is, and what it isn't

An adapter: LoRA weights on the T3 text-to-speech-token transformer, full-rank embeddings for the new tokens, and the extended tokenizer. You apply it on top of Chatterbox Multilingual v3, which you download from Resemble. We ship our delta, not a copy of their model.

Not merged weights, not a standalone model. Amharic only: Tigrinya and Ge'ez use the same script, but the adapter was not trained on them.

Amharic text front-end

amharic_text.py built the training labels, and the loader runs it on every input, so training and inference see the same text. One file, standard library only, same licence as the adapter, usable without the model. (One fix since training: a dotted abbreviation's trailing dot mid-sentence, as in … ዓ.ም. የአገሪቱ …, is no longer read as a full stop. Labels were built with the version whose SHA-256 is in training_config.json; the only effect on the model is one fewer spurious pause.)

It converts Ge'ez numerals, digits, decimals, percentages and clock times to words (1500 → አንድ ሺህ አምስት መቶ, 75% → ሰባ አምስት በመቶ, 3:30 → ሶስት ሰዓት ተኩል); expands about a hundred common abbreviations, keeping the inflected suffix (ዶ/ር → ዶክተር, ዓ.ም. → ዓመተ ምሕረት, መ/ቤቱ → መሥሪያ ቤቱ); collapses the four consonant families that are spelled several ways but pronounced the same (ሐ ኀ ኅ → ሀ, ሠ → ሰ, ዐ → አ, ፀ → ጸ); reduces punctuation to the five marks that change how you say something (። ፣ ፤ ? !); strips URLs, emoji and control characters. It's a subset of what we run in production.

from amharic_text import normalize, split_sentences
normalize("ዶ/ር አበበ በ2018 ዓ.ም በተደረገው ምርጫ 75% ድምፅ አገኙ።")
# 'ዶክተር አበበ በሁለት ሺህ አስራ ስምንት ኣመተ ምህረት በተደረገው ምርጫ ሰባ አምስት በመቶ ድምጽ አገኙ።'

Training data

Three sources. Every clip's filename starts with its corpus prefix; the assembled training directory was audited before training and the output is committed as is (audit/corpus_audit.txt). The adapter was trained from scratch on exactly that directory, starting from Resemble's stock v3 T3.

WaxalNLP comes as 48 kHz and Common Voice as 32/48 kHz MP3; both were resampled to 24 kHz. All Waxal Amharic clips that fit the trainer's 3 to 25 second window were used (190.93 h of roughly 191).

Licence: why CC-BY-SA-4.0

WaxalNLP is licensed under CC BY-SA 4.0, so the adapter is released under CC BY-SA 4.0. Credit to Digital Umuganda, the WaxalNLP contributors, and the Common Voice contributors. The base model is MIT; this licence covers what we add.

Architecture

Base: Chatterbox Multilingual v3, ResembleAI/chatterbox at revision 5bb1f6ee58e50c3b8d408bc82a6d3740c2db6e18, T3 file t3_mtl23ls_v3.safetensors. Pinned on purpose: the adapter only makes sense on that exact T3. v3 has Resemble's hallucination and speaker-similarity fixes over v2; S3Gen, the voice encoder and the tokenizer are the same as v2 and untouched.

Notes for anyone building on this:

No language token. The base tokenizer has [fr], [de] and so on; there is no [am] and we didn't add one. The adapter was trained on plain normalized text, and the loader tokenizes without a language prefix, without lower-casing or NFKD. language_id="am" on stock ChatterboxMultilingualTTS.generate raises ValueError; use the loader.

Alignment guard is off. Upstream enables its attention-alignment hallucination guard only when text_tokens_dict_size == 2454. With the extended vocabulary it's off, in training and at inference. The loader chunks by sentence instead, which handles the common failure (T3 stopping at the first sentence-final mark).

Train/inference parity. The shipped amharic_text.py differs from the file that built the training labels by later punctuation fixes. Both hashes and the reason are recorded in training_config.json as label_frontend_sha256, label_frontend_sha256_shipped and label_frontend_changed_since_training. Property checks over every training transcript are in audit/frontend_check.txt.

Evaluation

Held-out set: 100 clips from the same corpus, split before training by a seeded speaker-disjoint rule, so whole speakers are held out and none of their sentences appear in training. Checked independently of the trainer's own assertion: HELD OUT: eval ∩ train = ∅ at clip, speaker and sentence level; all eval stems are ih_/wxl_/cv_. (audit/holdout_verify.txt). Both models ran on the same clips with the same per-clip reference audio (the held-out speaker's own recording), through the same code: the released loader for the adapter, stock v3 with the same normalized text and no language tag.

Per-clip numbers: audit/eval/.

How CER is measured. We transcribe the generated audio with Meta's stock omniASR-CTC-3B, which we didn't train, and compare to the reference text after normalize_for_metric (collapses the homophone families so ሀ/ሐ/ኀ spellings don't count as errors, strips punctuation). Same ASR, same normalization, both models. The stock column is a floor: the base model can't read Fidel, so most of what it produces isn't Amharic. UTMOS and ECAPA involve no ASR. UTMOS was trained on English MOS ratings and both outputs are Amharic, so its near-tie says more about the metric than the models. Means are over clips that produced audio; failures are on their own row so they can't hide in an average.

Our own listening verdict: intelligible Amharic, not yet fully natural (read-aloud cadence, occasional flat prosody); the stock output is garbled and not Amharic.

Usage

pip install "chatterbox-tts @ git+https://github.com/resemble-ai/chatterbox@5de7a54aa4e5e2baadb0182dde554908b48b85c2" peft safetensors huggingface_hub torchaudio

from huggingface_hub import hf_hub_download
import importlib.util, torchaudio

# the loader + text front-end ship in this repo
spec = importlib.util.spec_from_file_location(
    "amharic_tts", hf_hub_download("gabar-tech/chatterbox-amharic", "amharic_tts.py"))
amharic_tts = importlib.util.module_from_spec(spec); spec.loader.exec_module(amharic_tts)

tts = amharic_tts.load_amharic_tts(device="cuda")   # downloads base v3 (pinned) + adapter
wav = tts.generate(
    "ሰላም! ይህ ከጽሑፍ በቀጥታ የተፈጠረ የአማርኛ ድምፅ ነው። ዛሬ ነሐሴ 11 ቀን 2018 ዓ.ም. ነው።",
    audio_prompt_path="reference.wav",     # ~10 s of the voice to clone, with consent
    temperature=0.6, cfg_weight=0.5)
torchaudio.save("out.wav", wav, tts.sr)      # 24 kHz, PerTh-watermarked
print(tts.normalize("ዛሬ ነሐሴ 11 ቀን 2018 ዓ.ም. ነው።"))   # what the model actually read

Or from a checkout: python amharic_tts.py "ሰላም ዓለም።" --ref reference.wav --out out.wav.

generate() normalizes the text, splits at sentence-final marks (T3 tends to stop at the first ። / ? / !), synthesizes each sentence against the reference and joins them. normalize=False / split_sentences=False turn those off. We ran the snippet above as written in a fresh virtualenv with only the packages listed, on NVIDIA RTX A6000, Ubuntu 22.04.5 LTS, python 3.11.10, torch 2.6.0+cu124, chatterbox-tts@5de7a54, peft 0.20.0, before publishing.

Watermarking

Chatterbox puts Resemble's PerTh watermark in every waveform it generates. We left that alone; nothing in this adapter touches S3Gen or the vocoder, where it happens. We ran the public resemble-perth detector over every demo clip from both models and the held-out eval outputs, with the natural reference recording as a negative control: present on all 115 checked files (detector confidence ≥ 0.5 on every generated file; the natural reference recording scores 0.0, so the detector is discriminating) (audit/watermark_verify.txt). If you build on this, leave it in. It's about the only provenance signal that exists for synthetic Amharic right now.

Intended use

• Amharic speech interfaces, audiobooks, education, accessibility, media production, with the consent of whoever's voice you clone.
• Research on low-resource TTS, script extension, Ethiopian language tech.

Out of scope

• Cloning someone's voice without their informed consent.
• Political persuasion, impersonating public figures, fraud (voice authentication included), harassment.
• Other languages. The adapter makes the base worse at its other languages.
• Anything safety-critical or broadcast without a human listening first.

Limitations

• Gemination (consonant length; contrastive in Amharic, unwritten in normal spelling) isn't marked. The model reads minimal pairs like አለ "said" / አለ "there is" from context and will get some wrong. Our internal models use a lexicon-based marker; we left it out so that what you type is what the model was trained on, with nothing private in between.
• Numbers, abbreviations and Ge'ez numerals are expanded by amharic_text.py. Skip it, or feed it things it doesn't handle (currency symbols, odd date formats, Latin words), and the model gets them raw.
• The first word or two of an utterance are sometimes slurred before the model settles; mid-sentence text is steadier. Sentences that open with a number phrase show it most.
• Ejectives (ቀ, ጠ, ጨ, ጰ, ጸ) are much better than nothing but not uniformly native.
• Very short inputs (a single word, "እሺ") can come out unstable. Put them in a sentence.
• Long passages: chunk at sentence boundaries. The loader does this for you.
• 193.63 hours, most of it read speech (Waxal and Common Voice are ASR corpora). Expect a neutral, read-aloud delivery; expressive and spontaneous speech is thin.
• Amharic only.

Risks and misuse

This model can clone a voice from roughly ten seconds of reference audio, so it can be misused for impersonation or fraud. Use it only with informed consent and human review. Outputs are watermarked by the shipped inference path. Every voice in the training data was recorded under a licence that permits this use.

Attribution

• Resemble AI, for Chatterbox Multilingual (MIT), the base model and the PerTh watermarker.
• Digital Umuganda and the WaxalNLP contributors, Amharic speech data (CC-BY-SA-4.0).
• Mozilla Common Voice contributors, Amharic speech data (CC0-1.0).

Citation

@misc{gabar2026chatterboxamharic,
  title  = {Chatterbox Amharic: a Fidel extension of Chatterbox Multilingual},
  author = {{Gabar Technologies}},
  year   = {2026},
  url    = {https://huggingface.co/gabar-tech/chatterbox-amharic}
}