Orpheus Australian-English TTS

A finetune of canopylabs/orpheus-tts-0.1-pretrained (Llama-3.2-3B backbone + SNAC 24 kHz neural codec) that speaks Australian-accented English in 8 named voices. Trained on public-domain Australian LibriVox audiobooks and dramatic readings, with mild DeepFilterNet denoising.

Voices

Pass one of these names as the speaker. (Names are random pseudonyms; the underlying public-domain readers are not identified.)

Samples

Generated with this model at temperature=0.6. Nothing was hand-picked by ear: each clip was transcribed back with Whisper large-v3-turbo and kept only if it matched the target sentence (31 of 32 first takes passed; one was re-rolled). Browse or download them in samples/, or open the demo Space.

“Good morning! It's a lovely day for a walk down by the harbour, isn't it?”

“The dance class starts at half past three, so we'd rather take the car than walk in this weather.”

“No worries, mate. I'll grab a coffee on the way and meet you at the station this afternoon.”

“It was a warm afternoon in the bush, and the road stretched on for miles ahead of us, past the grey gum trees and the dry creek bed.”

Install

pip install torch transformers soundfile
pip install snac            # neural audio codec used by Orpheus

If transformers fails to import with a Keras-3 error, run with USE_TF=0 USE_FLAX=0.

Quick start / deployment

A single self-contained inference function. Works on a CUDA GPU (bf16 ~7 GB VRAM); for CPU drop .to("cuda") and use float32.

import torch, soundfile as sf
from transformers import AutoModelForCausalLM, AutoTokenizer
from snac import SNAC

REPO = "ablmontazer/orpheus-australian-en-tts"
DEV  = "cuda"
# Orpheus control tokens (fixed)
SOH, EOT, EOH, SOAI, SOS, EOS, EOAI, PAD, BASE_A = 128259,128009,128260,128261,128257,128258,128262,128263,128266

tok   = AutoTokenizer.from_pretrained(REPO)
model = AutoModelForCausalLM.from_pretrained(REPO, torch_dtype=torch.bfloat16).to(DEV).eval()
snac  = SNAC.from_pretrained("hubertsiuzdak/snac_24khz").to(DEV).eval()

@torch.inference_mode()
def tts(voice: str, text: str, out_path: str = "out.wav", temperature: float = 0.6):
    # prompt format is  "<voice>: <text>"
    ids = [SOH] + tok.encode(f"{voice}: {text}", add_special_tokens=False) + [EOT, EOH, SOAI, SOS]
    gen = model.generate(torch.tensor([ids]).to(DEV), max_new_tokens=1400, do_sample=True,
                         temperature=temperature, top_p=0.95, eos_token_id=EOAI, pad_token_id=PAD)[0].tolist()
    audio = [t for t in gen[len(ids):] if BASE_A <= t < BASE_A + 7*4096]
    a = [t - BASE_A for t in audio]; n = len(a)//7
    c0, c1, c2 = [], [], []
    for i in range(n):                      # de-interleave 7 tokens/frame -> 3 SNAC codebooks
        f = a[i*7:i*7+7]
        c0.append(f[0]%4096)
        c1.append(f[1]%4096); c1.append(f[4]%4096)
        c2.append(f[2]%4096); c2.append(f[3]%4096); c2.append(f[5]%4096); c2.append(f[6]%4096)
    codes = [torch.tensor(c0)[None].to(DEV), torch.tensor(c1)[None].to(DEV), torch.tensor(c2)[None].to(DEV)]
    y = snac.decode(codes)[0, 0].cpu().numpy()
    sf.write(out_path, y, 24000)
    return out_path

# examples
tts("delta",  "Good morning! It's a lovely day for a walk down by the harbour.")
tts("flynn",  "No worries mate, I'll give you a hand with that.")

Available voices: flynn, angus, mason, cooper, delta, ruby, sienna, matilda.

Minimal HTTP server (optional)

from fastapi import FastAPI
from fastapi.responses import FileResponse
app = FastAPI()

@app.get("/tts")
def synth(voice: str, text: str):
    return FileResponse(tts(voice, text, "/tmp/tts.wav"), media_type="audio/wav")
# uvicorn app:app --host 0.0.0.0 --port 8000

Notes

• Sampling: temperature 0.5-0.7 is a good range; lower = steadier, higher = more expressive.
• Output is 24 kHz mono.
• One voice (the smallest) has less training data and is slightly less stable.

Training

• ~22 h total across the 8 voices (capped ~4 h/voice), 24 kHz, mild DeepFilterNet denoise.
• Sequence: [SOH] + text_ids("voice: text") + [EOT,EOH,SOAI,SOS] + snac_audio + [EOS,EOAI].
• Eval + early stopping (best checkpoint kept), cosine LR 5e-5, bf16, ~1 epoch effective.

Data & license

Source audio is public domain (LibriVox). Base model and code under Apache-2.0. Speaker identities are anonymised (pseudonymous voice names).