metadata
title: README
emoji: 🎙️
colorFrom: red
colorTo: yellow
sdk: static
pinned: false
Edmon Sahakyan · Armenian speech & dubbing
I build open models for Eastern Armenian (hy) speech: recognition, synthesis, translation, and end-to-end English → Armenian video dubbing for TV and drama.
Every card on this profile reports how the numbers were measured, and checkpoints that are still experimental are labelled as experimental.
Current models
| Model | Task | Status |
|---|---|---|
| speecht5_finetuned_voxpopuli_hy | Armenian TTS (SpeechT5, 2 speakers) | Main TTS checkpoint |
| TTS_NB_2 · TTS_NB_ONNX | Armenian TTS for training / ONNX | Active |
| speecht5_finetuned_hy | Armenian TTS (Common Voice) | Earlier checkpoint |
| whisper-small-hy | Armenian ASR | Experimental (validation WER ≈ 75%) |
| marian-finetuned-kde4-en-to-hy | English → Armenian MT (software domain) | Experimental |
Data: HyVoxPopuli (Armenian speech) · Demo: SpeechT5 Armenian TTS
Coming next: Armenian dubbing stack
From the armenian-video-dubbing project. Release names and dates are not fixed yet.
| Planned release | What it does | Licence target |
|---|---|---|
| Armenian voice-cloning TTS | Fine-tune of an open multilingual TTS on Common Voice hy-AM (CC0) and FLEURS hy_am (CC-BY-4.0) |
Commercial-friendly |
| Duration-aware EN → HY translation | Fits each Armenian line to the original timing (syllable budget) | Commercial-friendly |
| Armenian ASR adapter | Whisper LoRA for Armenian back-transcription and QA | Apache-2.0 / MIT base |
| Dubbing evaluation set | Public multi-speaker clips plus scoring notebook (WER, chrF++, COMET, SECS, accent leakage, LSE-C/D) | Per-source licences |
Each release will ship with a reproducible score from the ablation notebook, a licence statement covering every training source, and voice-cloning consent rules.
Principles
- Measured, not claimed. Metrics come from public test sets with a script anyone can rerun.
- Licences are explicit. Non-commercial components are marked and never used in a commercial release.
- Consent first. Voice cloning is only for speakers who agreed to it, and dubbed output is watermarked.
Contact
GitHub · LinkedIn · Bluesky · Open to collaboration on Armenian speech and language data.