| --- |
| license: cc-by-4.0 |
| language: |
| - am |
| pretty_name: OmniVoice Amharic Clean Speech — 500 Hours of Synthetic Call-Center Speech |
| tags: |
| - speech |
| - asr |
| - amharic |
| - ethiopia |
| - synthetic-speech |
| - call-center |
| - text-to-speech |
| size_categories: |
| - 100K<n<1M |
| task_categories: |
| - automatic-speech-recognition |
| - text-to-speech |
| --- |
| |
| # Amharic Corpus — Clean Speech |
|
|
| **146,359 clips · 408.5 hours · Amharic (አማርኛ) · 16 kHz · synthetic, studio-quality channel** |
|
|
| A large-scale **synthetic** Amharic speech corpus generated entirely by **OmniVoice**, the in-house Amharic text-to-speech system of [lab.et](https://lab.et) — the AI lab of Ethiopia. The audio simulates Ethiopian call-center conversations across banking, transport, telecom and everyday service scenarios, rendered with 62 distinct synthetic speaker identities (mixed genders). |
|
|
| ## Why this dataset exists |
|
|
| Amharic is spoken by 50+ million people but remains one of the world's most under-resourced languages for speech AI. Real Amharic call-center audio is scarce, expensive and privacy-sensitive. We built OmniVoice to close that gap: a high-fidelity Amharic TTS that lets us generate unlimited, perfectly-transcribed training data for automatic speech recognition and speech interfaces — without touching a single real customer's voice. |
|
|
| This release is the **clean channel** of the corpus: near-field, studio-quality speech. (A telephony-degraded channel variant exists but is not part of this release.) |
|
|
| ## Dataset details |
|
|
| | | | |
| |---|---| |
| | Clips | 146,359 | |
| | Total duration | ~408.5 hours | |
| | Language | Amharic (`am-ET`) | |
| | Sample rate | 16 kHz, mono, 16-bit WAV | |
| | Speakers | 62 synthetic identities (via voice cloning) | |
| | Generator | OmniVoice Amharic TTS (`omnivoice-amharic-finetuned-v3-clone`) | |
| | Domain | call-center dialogues — banking, transport, telecom, services | |
|
|
| ## Structure |
|
|
| Parquet shards in `data/`, each row: |
|
|
| | Field | Type | Description | |
| |---|---|---| |
| | `audio` | Audio | 16 kHz WAV bytes (embedded) | |
| | `text` | string | the exact transcript (what the TTS was asked to say) | |
| | `duration` | float | clip duration in seconds | |
| | `speaker_id` | string | synthetic speaker identity (e.g. `spk025`) | |
| | `gender` | string | `male` / `female` | |
| | `kind` | string | `base` (short utterance) / `long` (multi-turn) | |
| | `source_id` | string | internal generation id | |
|
|
| Load it with one line: |
|
|
| ```python |
| from datasets import load_dataset |
| ds = load_dataset("Lab-et/amharic-corpus", split="train") |
| ``` |
|
|
| ## Important limitations — read before use |
|
|
| 1. **This is synthetic speech.** It was machine-generated by an Amharic TTS system, not spoken by humans. It is excellent for ASR pre-training, domain adaptation and telephony research, but a model trained *only* on this data will not fully generalize to real human speech — combine it with human-recorded corpora (e.g. [Google WAXAL](https://huggingface.co/datasets/google/wax)) for best results. |
| 2. **Transcripts are exact by construction** (the TTS speaks the text), so there is no annotation noise in the label — but that also means the data carries no real-world disfluencies, accents or code-switching. |
| 3. **Speaker identities are synthetic** — cloned voices do not correspond to any real person. |
| 4. **No real customer data** was used at any stage: the dialogue text is synthetic, written by LLMs for training scenarios, then spoken by OmniVoice. |
|
|
| ## Ethics & compliance |
|
|
| - No real person's voice or personal data is contained in this dataset. |
| - Released under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) — attribution, any purpose, including commercial ASR training. |
| - Generated and published by an Ethiopian lab, for the benefit of Ethiopian-language AI. |
|
|
| ## About lab.et |
|
|
| [**lab.et**](https://lab.et) is the AI lab of Ethiopia — an artificial intelligence company in Addis Ababa building technology that works in the languages people here actually speak. |
|
|
| We exist because Ethiopian languages have been left behind by modern AI. Amharic alone is spoken by more than 50 million people, yet it remains one of the world's most under-resourced languages in speech and language technology. So we do the whole thing ourselves: we collect our own datasets, train our own models, and run our own infrastructure — no rented intelligence, no foreign data flows, and Ethiopian data stays in Ethiopia. |
|
|
| **Our products:** |
| - [**voice.et**](https://voice.et) — production speech AI: automatic speech recognition and natural text-to-speech voices for Amharic, Afaan Oromoo, Tigrinya and more, available as APIs and live voice products. |
| - [**yakal.et**](https://yakal.et) — the Ethiopian AI console: chat, images, video, voice and an AI site builder in Amharic, Afaan Oromoo and English, prepaid in Birr. |
|
|
| **Our services:** custom speech models tuned to an organization's domain (banking, healthcare, government services), integration into IVR systems, contact centers and telephony, and ethically sourced Ethiopian-language data for research. |
|
|
| **Open by default.** We believe the way to build an Ethiopian AI ecosystem is to fuel it — this corpus is our first major open release, and everything we publish will appear first at [huggingface.co/Lab-et](https://huggingface.co/Lab-et). |
|
|
| This corpus was generated by **OmniVoice**, our in-house Amharic text-to-speech system ([model here](https://huggingface.co/Lab-et/omnivoice-amharic)) — a good example of how our products build our data, and our data builds better products. |
|
|
| ## Citation |
|
|
| ```bibtex |
| @dataset{labet_amharic_corpus, |
| title = {Amharic Corpus — 400+ hours of synthetic call-center speech}, |
| author = {lab.et}, |
| year = {2026}, |
| url = {https://huggingface.co/datasets/Lab-et/amharic-corpus}, |
| note = {Generated with OmniVoice, the in-house Amharic TTS of lab.et} |
| } |
| ``` |
|
|
