Armenian SpeechT5 — training checkpoint (TTS_NB_2)

Recommended base for fine-tuning and experimentation. Successor to Edmon02/TTS_NB with an expanded tokenizer (vocab 158) and updated training on Common Voice 17 hy-AM + HyVoxPopuli.

Role	Training / research checkpoint
Vocab size	158 (vs 113 on speecht5_finetuned_voxpopuli_hy)
Used in	fine_tune_speecht5.py, notebooks 02, 04, 05

When to use which checkpoint

Goal	Model
Publish demos / share inference	speecht5_finetuned_voxpopuli_hy
Continue training / TRL / new data	TTS_NB_2 (this repo)
ONNX / edge deployment	TTS_NB_ONNX

Load for training

from transformers import SpeechT5Processor, SpeechT5ForTextToSpeech, SpeechT5Config

checkpoint = "Edmon02/TTS_NB_2"
processor = SpeechT5Processor.from_pretrained(checkpoint)
config = SpeechT5Config.from_pretrained(checkpoint)
model = SpeechT5ForTextToSpeech.from_pretrained(checkpoint)

Training data

• Edmon02/hyvoxpopuli
• mozilla-foundation/common_voice_17_0 / hy-AM (see project notebooks)

Lineage

Edmon02/speecht5_finetuned_voxpopuli_hy  (approx. branch)
  → Edmon02/TTS_NB
  → Edmon02/TTS_NB_2  ← you are here
  → Edmon02/TTS_NB_ONNX (export)

Limitations

• Internal codename (TTS_NB) — not descriptive; consider migrating to armenian-tts-speecht5-train
• Same domain limits as HyVoxPopuli (literary Armenian, 2 speakers)
• Requires microsoft/speecht5_hifigan at inference

Citation

@misc{armenian_speecht5_nb2_2024,
  author = {Avetisyan, Edmon},
  title = {Armenian SpeechT5 Training Checkpoint (TTS_NB_2)},
  year = {2024},
  howpublished = {\url{https://huggingface.co/Edmon02/TTS_NB_2}}
}

License

MIT