Views
No views yet
trust_remote_code porttrust_remote_code packaging of bosonai/higgs-audio-v3-tts-4b
that loads with plain 🤗 transformers (no SGLang). The weights are the original
checkpoint, copied unchanged; only a small modeling_*.py / configuration_*.py
pair and an auto_map were added.bosonai/higgs-audio-v2-tokenizer
(higgs_audio_v2_tokenizer), loaded automatically on first use.transformers >= 5.5.1import torch
2import torchaudio
3from transformers import AutoModelForCausalLM, AutoTokenizer
4
5repo = "multimodalart/higgs-audio-v3-tts-4b-transformers"
6tokenizer = AutoTokenizer.from_pretrained(repo)
7model = AutoModelForCausalLM.from_pretrained(
8 repo, trust_remote_code=True, dtype=torch.bfloat16
9).to("cuda").eval()
10
11# Zero-shot TTS
12wav = model.generate_speech("Hello, this is Higgs Audio running on transformers.", tokenizer)
13torchaudio.save("out.wav", wav.unsqueeze(0), model.config.sample_rate)
14
15# Voice cloning from a reference clip (+ optional transcript)
16ref, sr = torchaudio.load("reference.wav")
17wav = model.generate_speech(
18 "Now speaking in the cloned voice.",
19 tokenizer,
20 reference_audio=ref,
21 reference_sample_rate=sr,
22 reference_text="optional transcript of the reference clip",
23 temperature=0.7,
24 top_p=0.95,
25)
26torchaudio.save("clone.wav", wav.unsqueeze(0), model.config.sample_rate)generate_speech returns a mono 24 kHz waveform as a CPU float32 tensor [L].LICENSE.