Configuration Parsing Warning:In adapter_config.json: "peft.task_type" must be a string

Hilbert: OmniVoice LoRA voice

The voice of Hilbert, the sardonic, world-weary commentator of the My Weird Prompts podcast (https://myweirdprompts.com), baked into OmniVoice as a LoRA adapter, so it speaks in this voice without a reference clip.

Hilbert is a fictional character and the voice is synthetic: the training audio was generated by OmniVoice itself, zero-shot from the show's reference clip, then filtered and restored. No human voice was recorded for it.

How it was made (2026-11-14)

  1. Text: Hilbert's real lines from published episode transcripts, split into 6–45-word chunks, plus 40 short interjections ("Yeah.", "Right, okay.").
  2. Audio: OmniVoice (revision c5fdb5cc) zero-shot from the production reference clip, 2 takes per line.
  3. Filter: Whisper large-v3-turbo round trip (word-exact), duration check, ECAPA speaker similarity to the reference; best take per line. Hilbert: 184 of 230 lines kept (13.1 min); Sidon kept for 170.
  4. Restore + polish: Sidon v0.1 speech restoration; a light "warm" EQ/compression chain and loudnorm to -19 LUFS, 24 kHz mono.
  5. Train: OmniVoice's LoRA recipe (source at git 08be0b4): r 16, alpha 32, lr 1e-4, 600 steps on one L40S. Published checkpoint: checkpoint-600.

Use

import torch
from omnivoice import OmniVoice
from omnivoice.utils.lora import load_lora_adapter  # in OmniVoice's git source; not in the 0.2.1 wheel

model = OmniVoice.from_pretrained("k2-fsa/OmniVoice", device_map="cuda:0", dtype=torch.float16)
model = load_lora_adapter(model, "<local snapshot of this repo>")
audio = model.generate(text="I've read the press release so you don't have to.", language="English",
                       num_step=32, guidance_scale=2.0, denoise=True, postprocess_output=True)[0]

Samples (no reference clip)

Text Audio
Oh, wonderful. Another revolutionary breakthrough. I'll add it to the folder with the other four. samples/press.mp3
Can I just say, for the record, that I warned you about this? I have the email. It's dated. samples/warned.mp3
So the people who built one of the best open voice cloners now build the tools that catch voice clones. samples/sentence.mp3

Licence

CC-BY-NC-4.0, following the OmniVoice weights it adapts (CC-BY-NC, trained on Emilia). Non-commercial use only.

Recipe and scripts: the My Weird Prompts monorepo, planning/research/tts-landscape-2026-11/pilot-5-herman-lora-2026-11-14/.

Downloads last month
23
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for My-Weird-Prompts/Hilbert-OmniVoice-LoRA

Finetuned
Qwen/Qwen3-0.6B
Finetuned
k2-fsa/OmniVoice
Adapter
(60)
this model

Collections including My-Weird-Prompts/Hilbert-OmniVoice-LoRA