Spaces:
Running
on
Zero
Running
on
Zero
Demo Lina-speech (pardi-speech) on Spaces
Browse files- app.py +157 -0
- readme.md +11 -0
- requirements.txt +10 -0
app.py
ADDED
|
@@ -0,0 +1,157 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# app.py
|
| 2 |
+
import os
|
| 3 |
+
import gradio as gr
|
| 4 |
+
import numpy as np
|
| 5 |
+
import torch
|
| 6 |
+
import soundfile as sf
|
| 7 |
+
import spaces
|
| 8 |
+
|
| 9 |
+
from huggingface_hub import login
|
| 10 |
+
from pardi_speech import PardiSpeech, VelocityHeadSamplingParams # présent dans ce repo
|
| 11 |
+
# Les sous-modules requis se trouvent dans tts/ et codec/
|
| 12 |
+
|
| 13 |
+
MODEL_REPO_ID = os.environ.get("MODEL_REPO_ID", "theodorr/pardi-speech-enfr-forbidden")
|
| 14 |
+
|
| 15 |
+
# Auth HF (le secret est défini dans les Settings du Space)
|
| 16 |
+
HF_TOKEN = os.environ.get("HF_TOKEN")
|
| 17 |
+
if HF_TOKEN:
|
| 18 |
+
try:
|
| 19 |
+
login(token=HF_TOKEN)
|
| 20 |
+
print("✅ Logged to Hugging Face Hub.")
|
| 21 |
+
except Exception as e:
|
| 22 |
+
print("⚠️ HF login failed:", e)
|
| 23 |
+
|
| 24 |
+
# Chargement lazy pour ZeroGPU (et pour réduire le temps de cold-start)
|
| 25 |
+
_pardi = None
|
| 26 |
+
_sampling_rate = 24000
|
| 27 |
+
|
| 28 |
+
def _normalize_text(s: str, lang_hint: str = "fr") -> str:
|
| 29 |
+
"""Normalisation légère façon Whisper: lowercase + chiffres en lettres (si possible)."""
|
| 30 |
+
s = (s or "").strip().lower()
|
| 31 |
+
try:
|
| 32 |
+
import re
|
| 33 |
+
from num2words import num2words
|
| 34 |
+
def repl(m): return num2words(int(m.group()), lang=lang_hint)
|
| 35 |
+
s = re.sub(r"\d+", repl, s)
|
| 36 |
+
except Exception:
|
| 37 |
+
pass
|
| 38 |
+
return s
|
| 39 |
+
|
| 40 |
+
def _load_model(device: str = "cuda"):
|
| 41 |
+
global _pardi, _sampling_rate
|
| 42 |
+
if _pardi is None:
|
| 43 |
+
_pardi = PardiSpeech.from_pretrained(MODEL_REPO_ID, map_location=device)
|
| 44 |
+
_sampling_rate = getattr(_pardi, "sampling_rate", 24000)
|
| 45 |
+
print(f"✅ PardiSpeech loaded on {device} (sr={_sampling_rate}).")
|
| 46 |
+
return _pardi
|
| 47 |
+
|
| 48 |
+
def _to_mono_float32(arr: np.ndarray) -> np.ndarray:
|
| 49 |
+
arr = arr.astype(np.float32)
|
| 50 |
+
if arr.ndim == 2: # stereo -> mono
|
| 51 |
+
arr = arr.mean(axis=1)
|
| 52 |
+
return arr
|
| 53 |
+
|
| 54 |
+
@spaces.GPU(duration=120) # ZeroGPU: alloue un GPU pendant l'appel (noop ailleurs)
|
| 55 |
+
def synthesize(
|
| 56 |
+
text: str,
|
| 57 |
+
ref_audio, # tuple (sr, np.ndarray) ou chemin
|
| 58 |
+
ref_text: str,
|
| 59 |
+
steps: int,
|
| 60 |
+
cfg: float,
|
| 61 |
+
cfg_ref: float,
|
| 62 |
+
temperature: float,
|
| 63 |
+
max_seq_len: int,
|
| 64 |
+
seed: int,
|
| 65 |
+
lang_hint: str
|
| 66 |
+
):
|
| 67 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
| 68 |
+
torch.manual_seed(int(seed))
|
| 69 |
+
|
| 70 |
+
pardi = _load_model(device)
|
| 71 |
+
txt = _normalize_text(text, lang_hint=lang_hint)
|
| 72 |
+
|
| 73 |
+
# Prépare cache décodage
|
| 74 |
+
cache = pardi.tts.audio_decoder.init_cache(int(max_seq_len), device)
|
| 75 |
+
|
| 76 |
+
# Paramètres de sampling (cf. notebook inference & tes notes)
|
| 77 |
+
vel_params = VelocityHeadSamplingParams(
|
| 78 |
+
cfg_ref=float(cfg_ref),
|
| 79 |
+
cfg=float(cfg),
|
| 80 |
+
num_steps=int(steps),
|
| 81 |
+
temperature=float(temperature)
|
| 82 |
+
)
|
| 83 |
+
|
| 84 |
+
# Gestion du prefix (optionnel)
|
| 85 |
+
prefix = None
|
| 86 |
+
if ref_audio is not None:
|
| 87 |
+
# ref_audio peut être un chemin ou (sr, wav)
|
| 88 |
+
if isinstance(ref_audio, str):
|
| 89 |
+
wav, sr = sf.read(ref_audio)
|
| 90 |
+
else:
|
| 91 |
+
sr, wav = ref_audio
|
| 92 |
+
wav = _to_mono_float32(np.array(wav))
|
| 93 |
+
wav_t = torch.from_numpy(wav).to(device)
|
| 94 |
+
# Resample sur le sr attendu par le codec/ modèle
|
| 95 |
+
import torchaudio
|
| 96 |
+
if sr != pardi.sampling_rate:
|
| 97 |
+
wav_t = torchaudio.functional.resample(wav_t, sr, pardi.sampling_rate)
|
| 98 |
+
wav_t = wav_t.unsqueeze(0) # [1, T]
|
| 99 |
+
|
| 100 |
+
with torch.inference_mode():
|
| 101 |
+
# Encode prefix en tokens via PatchVAE (comme dans inference.ipynb)
|
| 102 |
+
prefix_tokens = pardi.patchvae.encode(wav_t) # [1, ...]
|
| 103 |
+
# ref_text est optionnel ; s’il est vide, on passe une chaîne vide
|
| 104 |
+
prefix = (ref_text or "", prefix_tokens[0])
|
| 105 |
+
|
| 106 |
+
# Synthèse
|
| 107 |
+
with torch.inference_mode():
|
| 108 |
+
wavs, _ = pardi.text_to_speech(
|
| 109 |
+
[txt],
|
| 110 |
+
prefix,
|
| 111 |
+
max_seq_len=int(max_seq_len),
|
| 112 |
+
velocity_head_sampling_params=vel_params,
|
| 113 |
+
cache=cache
|
| 114 |
+
)
|
| 115 |
+
wav = wavs[0].detach().cpu().numpy() # float32 [-1,1]
|
| 116 |
+
return (_sampling_rate, wav)
|
| 117 |
+
|
| 118 |
+
def build_demo():
|
| 119 |
+
with gr.Blocks(title="Lina‑speech / pardi‑speech Demo") as demo:
|
| 120 |
+
gr.Markdown(
|
| 121 |
+
"## Lina‑speech (pardi‑speech) – Démo TTS\n"
|
| 122 |
+
"Génère de l'audio à partir de texte, avec ou sans *prefix* (audio de référence).\n"
|
| 123 |
+
"Paramètres avancés: *num_steps*, *CFG*, *température*, *max_seq_len*, *seed*."
|
| 124 |
+
)
|
| 125 |
+
|
| 126 |
+
with gr.Row():
|
| 127 |
+
text = gr.Textbox(label="Texte à synthétiser", lines=4, placeholder="Tape ton texte ici…")
|
| 128 |
+
with gr.Accordion("Prefix (optionnel)", open=False):
|
| 129 |
+
ref_audio = gr.Audio(sources=["upload", "microphone"], type="numpy", label="Audio de référence")
|
| 130 |
+
ref_text = gr.Textbox(label="Texte du prefix (si connu)", placeholder="Transcription du prefix (optionnel)")
|
| 131 |
+
with gr.Accordion("Options avancées", open=False):
|
| 132 |
+
with gr.Row():
|
| 133 |
+
steps = gr.Slider(1, 50, value=10, step=1, label="num_steps")
|
| 134 |
+
cfg = gr.Slider(0.5, 3.0, value=1.4, step=0.05, label="CFG (guidance)")
|
| 135 |
+
cfg_ref = gr.Slider(0.5, 3.0, value=1.0, step=0.05, label="CFG (réf.)")
|
| 136 |
+
with gr.Row():
|
| 137 |
+
temperature = gr.Slider(0.1, 2.0, value=1.0, step=0.05, label="Température")
|
| 138 |
+
max_seq_len = gr.Slider(50, 1200, value=300, step=10, label="max_seq_len (tokens audio)")
|
| 139 |
+
seed = gr.Number(value=0, precision=0, label="Seed (reproductibilité)")
|
| 140 |
+
lang_hint = gr.Dropdown(choices=["fr", "en"], value="fr", label="Langue (normalisation)")
|
| 141 |
+
|
| 142 |
+
btn = gr.Button("Synthétiser")
|
| 143 |
+
out_audio = gr.Audio(label="Sortie audio", type="numpy")
|
| 144 |
+
|
| 145 |
+
# File d'attente pour GPU (gestion du débit)
|
| 146 |
+
demo.queue(default_concurrency_limit=1, max_size=32)
|
| 147 |
+
|
| 148 |
+
btn.click(
|
| 149 |
+
fn=synthesize,
|
| 150 |
+
inputs=[text, ref_audio, ref_text, steps, cfg, cfg_ref, temperature, max_seq_len, seed, lang_hint],
|
| 151 |
+
outputs=[out_audio]
|
| 152 |
+
)
|
| 153 |
+
return demo
|
| 154 |
+
|
| 155 |
+
if __name__ == "__main__":
|
| 156 |
+
demo = build_demo()
|
| 157 |
+
demo.launch()
|
readme.md
ADDED
|
@@ -0,0 +1,11 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Lina-speech (pardi-speech) — Demo Gradio
|
| 2 |
+
|
| 3 |
+
- Charge un checkpoint privé: `${MODEL_REPO_ID}` (par défaut `theodorr/pardi-speech-enfr-forbidden`)
|
| 4 |
+
- Nécessite un secret `HF_TOKEN` (Settings ▸ Secrets)
|
| 5 |
+
|
| 6 |
+
## Paramètres
|
| 7 |
+
- num_steps, CFG, CFG_ref, température, max_seq_len, seed
|
| 8 |
+
- Prefix optionnel: audio + texte (si disponible)
|
| 9 |
+
|
| 10 |
+
## Matériel
|
| 11 |
+
- ZeroGPU (PRO requis pour héberger) ou GPU T4/L4/L40S/A10G…
|
requirements.txt
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
gradio>=4.44.0
|
| 2 |
+
spaces>=0.20.0
|
| 3 |
+
huggingface_hub>=0.24.0
|
| 4 |
+
torch>=2.2.0
|
| 5 |
+
torchaudio>=2.2.0
|
| 6 |
+
numpy
|
| 7 |
+
soundfile
|
| 8 |
+
librosa
|
| 9 |
+
num2words
|
| 10 |
+
tqdm
|