Orpheus-3B Tajik TTS (Saidzoda Lab)

How to get access: please fill in the short form at saidzoda.com/access. We approve requests after we receive the form.

Full fine-tune of canopylabs/orpheus-3b-0.1-pretrained (Apache-2.0) for Tajik speech synthesis, using SNAC-tokenised Tajik speech (speaker 8, premium diarized bank). This is the run-8 (tg-e8) checkpoint, the best of a chained epoch-by-epoch fine-tune sweep (e1→e8, each epoch resuming from the previous epoch's final checkpoint).

Result

• Round-trip WER: 15.91% (CER 6.99%), measured by synthesizing 20 held-out Tajik utterances and transcribing them back with a Tajik-fine-tuned Whisper-large-v3 ASR model, beam=1, lowercase/strip-punctuation normalization (ё→е fold).
• This is the "spk8 epoch-8" champion referenced in ALL_TEST_RESULTS.md §7 — the best round-trip WER across the Orpheus Tajik bake-off sweep.

Training

• Method: full fine-tune (not LoRA), 1 epoch per stage, chained across 8 stages (tg-e1 … tg-e8), batch_size=1, lr=5e-5, pad_token=128263.
• Data: Tohirju/orpheus-tg-tokenised (SNAC-tokenised Tajik speech, speaker 8, faithful notebook port of the premium diarized bank).
• Base repo/commit: canopyai/Orpheus-TTS@e64661fe6d02c414fc77c53578c9d64082614861.
• Hardware: 1x H100 (Nvidia Brev).

How to use

This checkpoint is a standard Orpheus/Llama-architecture causal-LM checkpoint that emits SNAC audio codes, not raw waveform. To synthesize audio you need:

1. This model (text + reference-speaker prompt → SNAC token sequence).
2. The SNAC codec to decode those tokens to a 24kHz waveform: hubertsiuzdak/snac_24khz (MIT license).
3. The Orpheus-TTS inference code (or an equivalent SNAC-detokenizing wrapper) to drive generation and decode.

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("Tohirju/orpheus-tajik-tts")
tokenizer = AutoTokenizer.from_pretrained("Tohirju/orpheus-tajik-tts")
# then follow the Orpheus-TTS repo's inference path (prompt formatting + SNAC decode)

Files

• config.json, generation_config.json — model config
• model.safetensors — fp32 weights (~13.2 GB)
• tokenizer.json, tokenizer_config.json, special_tokens_map.json — tokenizer

Training-only artifacts (optimizer.pt, scheduler.pt, rng_state.pth, training_args.bin) are intentionally not included — they are not needed to serve the model and add ~26 GB of dead weight.

License

Apache-2.0 (inherited from the Orpheus-3B base model).

Caveats

• Round-trip WER is measured on 20 utterances synthesized from the training text pool (may overlap train) — a small, indicative eval, not a held-out generalization test.
• spk8 refers to the single reference speaker (speaker id 8 in the premium diarized bank) used for the SNAC-tokenised training data / eval prompts, not to a training epoch.

This repo is public but gated (manual approval), per Saidzoda Lab policy.