Spaces:
Running
on
Zero
Running
on
Zero
Update inference_cli.py
Browse files- inference_cli.py +118 -878
inference_cli.py
CHANGED
@@ -1,895 +1,135 @@
|
|
|
|
|
|
1 |
import argparse
|
2 |
-
import
|
3 |
-
import re
|
4 |
-
import tempfile
|
5 |
-
from pathlib import Path
|
6 |
-
import logging
|
7 |
-
import numpy as np
|
8 |
import soundfile as sf
|
9 |
-
import
|
10 |
-
import
|
11 |
-
import torchaudio
|
12 |
-
from tqdm import tqdm
|
13 |
-
from einops import rearrange
|
14 |
-
from pydub import AudioSegment, silence
|
15 |
-
from transformers import pipeline
|
16 |
-
from huggingface_hub import login
|
17 |
-
from cached_path import cached_path
|
18 |
-
import matplotlib.pyplot as plt # Needed for save_spectrogram
|
19 |
|
20 |
-
#
|
21 |
-
# !! Ensure these models are defined in your project's 'model' module !!
|
22 |
try:
|
23 |
-
|
24 |
-
|
25 |
-
|
26 |
-
|
27 |
-
|
28 |
-
|
29 |
-
|
30 |
-
|
31 |
-
|
32 |
-
|
33 |
-
|
34 |
-
|
35 |
-
|
36 |
-
|
37 |
-
|
38 |
-
|
39 |
-
# ---
|
40 |
-
|
41 |
-
|
42 |
-
|
43 |
-
|
44 |
-
# ---
|
45 |
-
|
46 |
-
#
|
47 |
-
|
48 |
-
"""
|
49 |
-
|
50 |
-
|
51 |
-
|
52 |
-
|
53 |
-
|
54 |
-
|
55 |
-
|
56 |
-
|
57 |
-
|
58 |
-
|
59 |
-
|
60 |
-
|
61 |
-
|
62 |
-
|
63 |
-
|
64 |
-
|
65 |
-
|
66 |
-
|
67 |
-
|
68 |
-
|
69 |
-
|
70 |
-
|
71 |
-
|
72 |
-
|
73 |
-
"
|
74 |
-
print("
|
75 |
-
|
76 |
-
|
77 |
-
|
78 |
-
trimmed_aseg = aseg[start_trim:duration-end_trim]
|
79 |
-
print(f"Removed {start_trim}ms from start, {end_trim}ms from end.")
|
80 |
-
return trimmed_aseg
|
81 |
-
|
82 |
-
# Function to save spectrogram (from app.py)
|
83 |
-
def save_spectrogram(spectrogram, file_path):
|
84 |
-
"""Saves a spectrogram visualization to a file."""
|
85 |
-
if spectrogram is None:
|
86 |
-
print("Spectrogram data is None, cannot save.")
|
87 |
-
return
|
88 |
-
try:
|
89 |
-
print(f"Saving spectrogram to {file_path}...")
|
90 |
-
plt.figure(figsize=(10, 4))
|
91 |
-
plt.imshow(spectrogram, aspect='auto', origin='lower', cmap='viridis')
|
92 |
-
plt.colorbar(label='Mel power')
|
93 |
-
plt.xlabel('Frames')
|
94 |
-
plt.ylabel('Mel bins')
|
95 |
-
plt.title('Generated Mel Spectrogram')
|
96 |
-
plt.tight_layout()
|
97 |
-
plt.savefig(file_path)
|
98 |
-
plt.close() # Close the figure to free memory
|
99 |
-
print("Spectrogram saved.")
|
100 |
-
except Exception as e:
|
101 |
-
print(f"Error saving spectrogram: {e}")
|
102 |
-
|
103 |
-
# Helper function to load checkpoint (from app.py, slightly modified for CLI)
|
104 |
-
def load_checkpoint(model, ckpt_path, device, use_ema=False):
|
105 |
-
"""Loads model weights from a checkpoint file (.pt or .safetensors)."""
|
106 |
-
print(f"Loading checkpoint from {ckpt_path}...")
|
107 |
try:
|
108 |
-
|
109 |
-
|
110 |
-
|
111 |
-
|
112 |
-
|
113 |
-
|
114 |
-
|
115 |
-
|
116 |
-
|
117 |
-
|
118 |
-
|
119 |
-
|
120 |
-
|
121 |
-
# Handle EMA weights
|
122 |
-
ema_key_prefix = "ema_model." # Adjust if your EMA keys have a different prefix
|
123 |
-
final_state_dict = {}
|
124 |
-
has_ema = any(k.startswith(ema_key_prefix) for k in state_dict.keys())
|
125 |
-
|
126 |
-
if use_ema:
|
127 |
-
if has_ema:
|
128 |
-
print("Attempting to load EMA weights.")
|
129 |
-
ema_state_dict = {k[len(ema_key_prefix):]: v for k, v in state_dict.items() if k.startswith(ema_key_prefix)}
|
130 |
-
if ema_state_dict:
|
131 |
-
final_state_dict = ema_state_dict
|
132 |
-
print("Using EMA weights.")
|
133 |
-
else:
|
134 |
-
# This case shouldn't happen if has_ema is true, but as a safeguard:
|
135 |
-
print("Warning: EMA weights requested but none found starting with prefix. Using regular weights.")
|
136 |
-
final_state_dict = {k: v for k, v in state_dict.items() if not k.startswith(ema_key_prefix)}
|
137 |
-
else:
|
138 |
-
print("Warning: EMA weights requested but no keys found with EMA prefix. Using regular weights.")
|
139 |
-
final_state_dict = state_dict # Use the original dict if no EMA keys exist
|
140 |
-
else:
|
141 |
-
print("Loading non-EMA weights.")
|
142 |
-
# Filter out EMA weights if they exist and we explicitly don't want them
|
143 |
-
final_state_dict = {k: v for k, v in state_dict.items() if not k.startswith(ema_key_prefix)}
|
144 |
-
|
145 |
-
|
146 |
-
# Load into model, handling potential 'module.' prefix from DDP
|
147 |
-
model_state_dict = model.state_dict()
|
148 |
-
processed_state_dict = {}
|
149 |
-
for k, v in final_state_dict.items():
|
150 |
-
if k.startswith("module."):
|
151 |
-
k_proc = k[len("module."):]
|
152 |
-
else:
|
153 |
-
k_proc = k
|
154 |
-
|
155 |
-
if k_proc in model_state_dict:
|
156 |
-
if model_state_dict[k_proc].shape == v.shape:
|
157 |
-
processed_state_dict[k_proc] = v
|
158 |
-
else:
|
159 |
-
print(f"Warning: Shape mismatch for key {k_proc}. Checkpoint: {v.shape}, Model: {model_state_dict[k_proc].shape}. Skipping.")
|
160 |
-
# else: # Optional: Log unexpected keys
|
161 |
-
# print(f"Warning: Key {k_proc} from checkpoint not found in model. Skipping.")
|
162 |
-
|
163 |
-
missing_keys, unexpected_keys = model.load_state_dict(processed_state_dict, strict=False)
|
164 |
-
|
165 |
-
if missing_keys:
|
166 |
-
print(f"Warning: Missing keys in model not found in checkpoint: {missing_keys}")
|
167 |
-
if unexpected_keys:
|
168 |
-
# This should ideally be empty if we filter correctly, but good to check.
|
169 |
-
print(f"Warning: Unexpected keys (should not happen with filtering): {unexpected_keys}")
|
170 |
-
|
171 |
-
print(f"Checkpoint loaded successfully from {ckpt_path}")
|
172 |
-
|
173 |
-
except FileNotFoundError:
|
174 |
-
print(f"Error: Checkpoint file not found at {ckpt_path}")
|
175 |
-
raise
|
176 |
-
except Exception as e:
|
177 |
-
print(f"Error loading checkpoint from {ckpt_path}: {e}")
|
178 |
-
raise # Re-raise the exception
|
179 |
-
|
180 |
-
model.eval()
|
181 |
-
return model.to(device)
|
182 |
-
|
183 |
-
# Primary model loading function (from app.py)
|
184 |
-
def load_custom(model_cls, model_cfg, ckpt_path: str, vocab_size: int, device='cpu', use_ema=True):
|
185 |
-
"""Loads a custom TTS model (DiT or UNetT) with specified config and checkpoint."""
|
186 |
-
ckpt_path = ckpt_path.strip()
|
187 |
-
|
188 |
-
if ckpt_path.startswith("hf://"):
|
189 |
-
print(f"Downloading checkpoint from Hugging Face Hub: {ckpt_path}")
|
190 |
-
try:
|
191 |
-
ckpt_path = str(cached_path(ckpt_path))
|
192 |
-
print(f"Checkpoint downloaded to: {ckpt_path}")
|
193 |
-
except Exception as e:
|
194 |
-
print(f"Error downloading checkpoint {ckpt_path}: {e}")
|
195 |
-
raise
|
196 |
-
|
197 |
-
if not Path(ckpt_path).exists():
|
198 |
-
raise FileNotFoundError(f"Checkpoint file not found: {ckpt_path}")
|
199 |
-
|
200 |
-
# Ensure necessary config keys are present (add defaults if missing)
|
201 |
-
if 'mel_dim' not in model_cfg:
|
202 |
-
model_cfg['mel_dim'] = 100 # Default mel channels
|
203 |
-
print(f"Warning: 'mel_dim' not in model_cfg, defaulting to {model_cfg['mel_dim']}")
|
204 |
-
if 'text_num_embeds' not in model_cfg:
|
205 |
-
model_cfg['text_num_embeds'] = vocab_size
|
206 |
-
print(f"Setting 'text_num_embeds' in model_cfg to vocab size: {vocab_size}")
|
207 |
-
|
208 |
-
print(f"Instantiating model: {model_cls.__name__} with config: {model_cfg}")
|
209 |
try:
|
210 |
-
|
211 |
-
|
212 |
-
|
213 |
-
|
214 |
-
|
215 |
-
|
216 |
-
|
217 |
-
|
218 |
-
|
219 |
-
|
220 |
-
|
221 |
-
# Text chunking function (from app.py)
|
222 |
-
def chunk_text(text, max_chars):
|
223 |
-
"""
|
224 |
-
Splits the input text into chunks based on punctuation and length limits.
|
225 |
-
(Copied from previous answer, assumed correct)
|
226 |
-
"""
|
227 |
-
if not isinstance(text, str):
|
228 |
-
print("Warning: Input to chunk_text is not a string. Returning empty list.")
|
229 |
-
return []
|
230 |
-
|
231 |
-
if max_chars > 135:
|
232 |
-
print(f"Warning: Calculated max_chars ({max_chars}) > 135. Capping at 135.")
|
233 |
-
max_chars = 135
|
234 |
-
if max_chars < 50:
|
235 |
-
print(f"Warning: Calculated max_chars ({max_chars}) < 50. Setting to 50.")
|
236 |
-
max_chars = 50
|
237 |
-
|
238 |
-
split_after_space_chars = max_chars + int(max_chars * 0.33)
|
239 |
-
chunks = []
|
240 |
-
current_chunk = ""
|
241 |
-
|
242 |
-
# Split the text into sentences based on punctuation followed by whitespace
|
243 |
-
sentences = re.split(r"(?<=[;:,.!?])\s+|(?<=[;:,。!?])\s*", text) # Added \s* after CJK punc
|
244 |
-
|
245 |
-
for sentence in sentences:
|
246 |
-
sentence = sentence.strip()
|
247 |
-
if not sentence:
|
248 |
-
continue
|
249 |
-
|
250 |
-
# Estimate potential length increase due to space
|
251 |
-
estimated_len = len(current_chunk) + len(sentence) + (1 if current_chunk else 0)
|
252 |
-
|
253 |
-
if estimated_len <= max_chars:
|
254 |
-
current_chunk += (" " + sentence) if current_chunk else sentence
|
255 |
-
else:
|
256 |
-
# Process the current_chunk if adding the new sentence exceeds max_chars
|
257 |
-
while len(current_chunk) > split_after_space_chars:
|
258 |
-
split_index = current_chunk.rfind(" ", 0, split_after_space_chars)
|
259 |
-
if split_index == -1: split_index = split_after_space_chars
|
260 |
-
chunks.append(current_chunk[:split_index].strip())
|
261 |
-
current_chunk = current_chunk[split_index:].strip()
|
262 |
-
|
263 |
-
if current_chunk:
|
264 |
-
chunks.append(current_chunk)
|
265 |
-
|
266 |
-
# Start new chunk, handle if sentence itself is too long
|
267 |
-
while len(sentence) > split_after_space_chars:
|
268 |
-
split_index = sentence.rfind(" ", 0, split_after_space_chars)
|
269 |
-
if split_index == -1: split_index = split_after_space_chars
|
270 |
-
chunks.append(sentence[:split_index].strip())
|
271 |
-
sentence = sentence[split_index:].strip()
|
272 |
-
current_chunk = sentence
|
273 |
-
|
274 |
-
# Handle the last chunk
|
275 |
-
while len(current_chunk) > split_after_space_chars:
|
276 |
-
split_index = current_chunk.rfind(" ", 0, split_after_space_chars)
|
277 |
-
if split_index == -1: split_index = split_after_space_chars
|
278 |
-
chunks.append(current_chunk[:split_index].strip())
|
279 |
-
current_chunk = current_chunk[split_index:].strip()
|
280 |
-
|
281 |
-
if current_chunk:
|
282 |
-
chunks.append(current_chunk.strip())
|
283 |
-
|
284 |
-
return [c for c in chunks if c] # Filter empty chunks
|
285 |
-
|
286 |
-
|
287 |
-
# Text to IPA function (from app.py)
|
288 |
-
def text_to_ipa(text, language):
|
289 |
-
"""Converts text to IPA using phonemizer with espeak backend."""
|
290 |
-
if not isinstance(text, str) or not text.strip():
|
291 |
-
print(f"Warning: Invalid input text for IPA conversion: {text}")
|
292 |
-
return "" # Return empty string for invalid input
|
293 |
-
try:
|
294 |
-
# Ensure phonemizer is installed: pip install phonemizer
|
295 |
-
# Ensure espeak-ng is installed: sudo apt-get install espeak-ng (or equivalent)
|
296 |
-
ipa_text = phonemize(
|
297 |
-
text,
|
298 |
-
language=language,
|
299 |
-
backend='espeak',
|
300 |
-
strip=False, # Keep punctuation
|
301 |
-
preserve_punctuation=True,
|
302 |
-
with_stress=True,
|
303 |
-
language_switch='remove-flags', # Use this instead of regex removal
|
304 |
-
njobs=1 # Set njobs=1 for potentially better stability/simpler debugging
|
305 |
)
|
306 |
-
|
307 |
-
ipa_text = re.sub(r'tʃˈaɪniːzlˈe̞tə', '', ipa_text)
|
308 |
-
ipa_text = re.sub(r'tʃˈaɪniːzɭˈetə', '', ipa_text)
|
309 |
-
ipa_text = re.sub(r'dʒˈapəniːzlˈe̞tə', '', ipa_text)
|
310 |
-
ipa_text = re.sub(r'dʒˈapəniːzɭˈetə', '', ipa_text)
|
311 |
-
|
312 |
-
ipa_text = ipa_text.strip()
|
313 |
-
# Replace multiple spaces with single space
|
314 |
-
ipa_text = re.sub(r'\s+', ' ', ipa_text)
|
315 |
|
316 |
-
print(f"Text: '{text}' | Lang: {language} | IPA: '{ipa_text}'")
|
317 |
-
return ipa_text
|
318 |
-
except ImportError:
|
319 |
-
print("Error: 'phonemizer' library not found. Please install it: pip install phonemizer")
|
320 |
-
raise
|
321 |
except Exception as e:
|
322 |
-
|
323 |
-
|
324 |
-
|
325 |
-
|
326 |
-
else:
|
327 |
-
print(f"Error phonemizing text: '{text}' with language '{language}'. Error: {e}")
|
328 |
-
# Decide how to handle error
|
329 |
-
raise ValueError(f"Phonemization failed for '{text}' ({language})") from e
|
330 |
-
|
331 |
-
|
332 |
-
# --- End of functions from app.py ---
|
333 |
-
|
334 |
-
# --- Argument Parser Setup ---
|
335 |
-
# (Parser definition remains the same as previous refactored version)
|
336 |
-
parser = argparse.ArgumentParser(
|
337 |
-
prog="python3 inference-cli.py",
|
338 |
-
description="Commandline interface for F5/E2 TTS.",
|
339 |
-
)
|
340 |
-
parser.add_argument(
|
341 |
-
"-c", "--config", type=str, default="inference-cli.toml",
|
342 |
-
help="Path to configuration file (TOML format). Default: inference-cli.toml"
|
343 |
-
)
|
344 |
-
# --- Arguments overriding config or providing inputs ---
|
345 |
-
parser.add_argument( "--ckpt_path", type=str, default=None, help="Path or Hub ID (hf://...) to the TTS model checkpoint (.pt/.safetensors). Overrides config.")
|
346 |
-
parser.add_argument( "--ref_audio", type=str, default=None, help="Path to the reference audio file (<10s recommended). Overrides config.")
|
347 |
-
parser.add_argument( "--ref_text", type=str, default=None, help="Reference text. If omitted, Whisper transcription is used. Overrides config.")
|
348 |
-
parser.add_argument( "--gen_text", type=str, default=None, help="Text to synthesize. Overrides config.")
|
349 |
-
parser.add_argument( "--gen_file", type=str, default=None, help="File containing text to synthesize (overrides --gen_text and config).")
|
350 |
-
parser.add_argument( "--output_dir", type=str, default=None, help="Directory to save output audio and spectrogram. Overrides config.")
|
351 |
-
parser.add_argument( "--output_name", type=str, default="out", help="Base name for output files (e.g., 'my_speech' -> my_speech.wav, my_speech.png). Default: out.")
|
352 |
-
# --- Parameter Arguments ---
|
353 |
-
parser.add_argument( "--ref_language", type=str, default=None, help="Language code for reference text phonemization (e.g., 'en-us', 'pl', 'de'). Overrides config.")
|
354 |
-
parser.add_argument( "--language", type=str, default=None, help="Language code for generated text phonemization (e.g., 'en-us', 'pl', 'de'). Overrides config.")
|
355 |
-
parser.add_argument( "--speed", type=float, default=None, help="Speech speed multiplier. Overrides config.")
|
356 |
-
parser.add_argument( "--nfe", type=int, default=None, help="Number of function evaluations (sampling steps). Overrides config.")
|
357 |
-
parser.add_argument( "--cfg", type=float, default=None, help="Classifier-Free Guidance strength. Overrides config.")
|
358 |
-
parser.add_argument( "--sway", type=float, default=None, help="Sway sampling coefficient. Overrides config.")
|
359 |
-
parser.add_argument( "--cross_fade", type=float, default=None, help="Cross-fade duration between batches (seconds). Overrides config.")
|
360 |
-
parser.add_argument( "--remove_silence", action=argparse.BooleanOptionalAction, default=None, help="Enable/disable final silence removal. Overrides config.")
|
361 |
-
parser.add_argument( "--hf_token", type=str, default=None, help="Hugging Face API token (for downloading private models/checkpoints).")
|
362 |
-
parser.add_argument( "--tokenizer_path", type=str, default=None, help="Path to the tokenizer.json file. Overrides config.")
|
363 |
-
parser.add_argument( "--device", type=str, default=None, help="Device to use ('cuda', 'cpu', 'mps'). Auto-detects if not set.")
|
364 |
-
parser.add_argument( "--dtype", type=str, default=None, help="Data type ('float16', 'bfloat16', 'float32'). Auto-selects if not set.")
|
365 |
-
|
366 |
-
args = parser.parse_args()
|
367 |
|
368 |
-
# ---
|
369 |
-
config = {}
|
370 |
-
if Path(args.config).exists():
|
371 |
try:
|
372 |
-
|
373 |
-
|
374 |
-
|
375 |
-
|
376 |
-
|
377 |
-
|
378 |
-
|
379 |
-
|
380 |
-
|
381 |
-
#
|
382 |
-
|
383 |
-
|
384 |
-
|
385 |
-
|
386 |
-
|
387 |
-
|
388 |
-
|
389 |
-
|
390 |
-
|
391 |
-
language = args.language or config.get("language", "en-us")
|
392 |
-
speed = args.speed if args.speed is not None else config.get("speed", 1.0)
|
393 |
-
nfe_step = args.nfe if args.nfe is not None else config.get("nfe", 32)
|
394 |
-
cfg_strength = args.cfg if args.cfg is not None else config.get("cfg", 2.0)
|
395 |
-
sway_sampling_coef = args.sway if args.sway is not None else config.get("sway", -1.0)
|
396 |
-
cross_fade_duration = args.cross_fade if args.cross_fade is not None else config.get("cross_fade", 0.15)
|
397 |
-
remove_silence_flag = args.remove_silence if args.remove_silence is not None else config.get("remove_silence", False)
|
398 |
-
hf_token = args.hf_token or config.get("hf_token")
|
399 |
-
tokenizer_path = args.tokenizer_path or config.get("tokenizer_path", "data/Emilia_ZH_EN_pinyin/tokenizer.json")
|
400 |
-
|
401 |
-
|
402 |
-
# --- Validate Required Arguments ---
|
403 |
-
if not ckpt_path: raise ValueError("Missing required argument/config: --ckpt_path")
|
404 |
-
if not ref_audio_path: raise ValueError("Missing required argument/config: --ref_audio")
|
405 |
-
if not gen_text and not gen_file: raise ValueError("Missing required argument/config: --gen_text or --gen_file")
|
406 |
-
|
407 |
-
# --- Read gen_text from file if provided ---
|
408 |
-
if gen_file:
|
409 |
-
try:
|
410 |
-
with codecs.open(gen_file, "r", "utf-8") as f: gen_text = f.read()
|
411 |
-
print(f"Loaded generation text from {gen_file}")
|
412 |
-
except Exception as e: raise ValueError(f"Error reading generation text file {gen_file}: {e}")
|
413 |
-
|
414 |
-
# --- Setup Device and Dtype ---
|
415 |
-
# (Device/Dtype setup remains the same)
|
416 |
-
cli_device = args.device or config.get("device")
|
417 |
-
if cli_device:
|
418 |
-
device = torch.device(cli_device)
|
419 |
-
else:
|
420 |
-
device = torch.device("cuda" if torch.cuda.is_available() else "mps" if torch.backends.mps.is_available() else "cpu")
|
421 |
-
|
422 |
-
cli_dtype = args.dtype or config.get("dtype")
|
423 |
-
if cli_dtype:
|
424 |
-
dtype_map = {"float16": torch.float16, "bfloat16": torch.bfloat16, "float32": torch.float32}
|
425 |
-
if cli_dtype in dtype_map: dtype = dtype_map[cli_dtype]
|
426 |
-
else: raise ValueError(f"Unsupported dtype: {cli_dtype}")
|
427 |
-
else:
|
428 |
-
if device.type == "cuda": dtype = torch.float16
|
429 |
-
elif device.type == "cpu" and hasattr(torch.backends, 'cpu') and torch.backends.cpu.supports_bfloat16: dtype = torch.bfloat16
|
430 |
-
else: dtype = torch.float32
|
431 |
|
432 |
-
print(f"Using device: {device}, dtype: {dtype}")
|
433 |
-
|
434 |
-
# --- Hugging Face Login ---
|
435 |
-
if hf_token:
|
436 |
-
print("Logging in to Hugging Face Hub...")
|
437 |
-
try:
|
438 |
-
login(token=hf_token)
|
439 |
-
print("Logged in successfully.")
|
440 |
except Exception as e:
|
441 |
-
print(f"
|
442 |
-
|
443 |
-
|
444 |
-
# --- Create Output Directory ---
|
445 |
-
output_dir.mkdir(parents=True, exist_ok=True)
|
446 |
-
wave_path = output_dir / f"{output_name}.wav"
|
447 |
-
spectrogram_path = output_dir / f"{output_name}.png"
|
448 |
|
449 |
-
#
|
450 |
-
|
451 |
-
try:
|
452 |
-
if
|
453 |
-
|
454 |
-
|
455 |
-
|
456 |
-
print(f"Tokenizer loaded successfully. Vocab size: {vocab_size}")
|
457 |
-
except Exception as e:
|
458 |
-
raise ValueError(f"Error loading tokenizer from {tokenizer_path}: {e}")
|
459 |
|
460 |
-
print("Loading Vocoder...")
|
461 |
-
# Pass device to load_vocoder
|
462 |
-
vocos = load_vocoder(device=device) # Already includes .to(device).eval()
|
463 |
-
|
464 |
-
print("Loading ASR Model (Whisper)...")
|
465 |
-
try:
|
466 |
-
whisper_dtype = torch.float16 if device.type == 'cuda' else torch.float32
|
467 |
-
# Reduce default batch_size for Whisper CLI use
|
468 |
-
pipe = pipeline(
|
469 |
-
"automatic-speech-recognition",
|
470 |
-
model="openai/whisper-large-v3-turbo",
|
471 |
-
torch_dtype=whisper_dtype,
|
472 |
-
device=device,
|
473 |
-
model_kwargs={"attn_implementation": "sdpa"} # Use SDPA if available
|
474 |
-
)
|
475 |
-
print("Whisper model loaded.")
|
476 |
-
except Exception as e:
|
477 |
-
print(f"Warning: Could not load Whisper ASR model: {e}. Transcription will not be available.")
|
478 |
-
pipe = None
|
479 |
-
|
480 |
-
print("Loading TTS Model...")
|
481 |
-
# --- Determine Model Class and Config ---
|
482 |
-
# Example configs (ensure they match your actual model requirements)
|
483 |
-
F5TTS_model_cfg = dict(dim=1024, depth=22, heads=16, ff_mult=2, text_dim=512, conv_layers=4)
|
484 |
-
E2TTS_model_cfg = dict(dim=1024, depth=24, heads=16, ff_mult=4) # Add mel_dim/text_num_embeds if needed by class
|
485 |
-
|
486 |
-
# Heuristic to determine model class (improve if needed)
|
487 |
-
if "E2TTS" in ckpt_path or "UNetT" in ckpt_path:
|
488 |
-
model_cls = UNetT
|
489 |
-
model_cfg = E2TTS_model_cfg
|
490 |
-
print(f"Assuming E2-TTS (UNetT) architecture for {ckpt_path}.")
|
491 |
-
elif "F5TTS" in ckpt_path or "DiT" in ckpt_path:
|
492 |
-
model_cls = DiT
|
493 |
-
model_cfg = F5TTS_model_cfg
|
494 |
-
print(f"Assuming F5-TTS (DiT) architecture for {ckpt_path}.")
|
495 |
-
else:
|
496 |
-
# Default or raise error if model type cannot be inferred
|
497 |
-
print(f"Warning: Cannot infer model type from '{ckpt_path}'. Defaulting to DiT/F5TTS.")
|
498 |
-
model_cls = DiT
|
499 |
-
model_cfg = F5TTS_model_cfg
|
500 |
-
|
501 |
-
|
502 |
-
try:
|
503 |
-
# Pass vocab_size needed by load_custom
|
504 |
-
ema_model = load_custom(model_cls, model_cfg, ckpt_path, vocab_size=vocab_size, device=device, use_ema=True)
|
505 |
-
# Ensure model is using the target runtime dtype
|
506 |
-
ema_model = ema_model.to(dtype=dtype)
|
507 |
-
print(f"TTS Model loaded successfully ({model_cls.__name__}).")
|
508 |
-
except Exception as e:
|
509 |
-
print(f"Critical Error: Failed to load TTS model from {ckpt_path}: {e}")
|
510 |
-
raise
|
511 |
-
|
512 |
-
# --- Settings from app.py ---
|
513 |
-
target_sample_rate = 24000
|
514 |
-
n_mel_channels = model_cfg.get('mel_dim', 100) # Use mel_dim from config if available
|
515 |
-
hop_length = 256
|
516 |
-
target_rms = 0.1
|
517 |
-
|
518 |
-
# --- Main Inference Logic ---
|
519 |
-
|
520 |
-
def infer_batch(ref_audio_tuple, ref_text_ipa, gen_text_ipa_batches,
|
521 |
-
ema_model, vocos, tokenizer,
|
522 |
-
remove_silence_post, cross_fade_duration,
|
523 |
-
nfe_step, cfg_strength, sway_sampling_coef, speed,
|
524 |
-
target_sample_rate, hop_length, target_rms, device, dtype):
|
525 |
-
"""
|
526 |
-
Generates audio batches based on reference and text inputs.
|
527 |
-
(Function body remains the same as previous refactored version)
|
528 |
-
"""
|
529 |
-
audio, sr = ref_audio_tuple
|
530 |
-
audio = audio.to(device, dtype=dtype)
|
531 |
-
|
532 |
-
# Preprocess reference audio (resample, RMS norm)
|
533 |
-
if audio.shape[0] > 1: audio = torch.mean(audio, dim=0, keepdim=True)
|
534 |
-
current_rms = torch.sqrt(torch.mean(torch.square(audio)))
|
535 |
-
rms_applied_factor = 1.0 # Track scaling factor applied to ref
|
536 |
-
if current_rms < target_rms and current_rms > 1e-5: # Add safety check for near-silent audio
|
537 |
-
print(f"Reference audio RMS ({current_rms:.3f}) below target ({target_rms}). Normalizing.")
|
538 |
-
rms_applied_factor = target_rms / current_rms
|
539 |
-
audio = audio * rms_applied_factor
|
540 |
-
elif current_rms <= 1e-5:
|
541 |
-
print("Warning: Reference audio is near silent. Skipping RMS normalization.")
|
542 |
-
else:
|
543 |
-
print(f"Reference audio RMS ({current_rms:.3f}) >= target ({target_rms}). No normalization.")
|
544 |
-
|
545 |
-
if sr != target_sample_rate:
|
546 |
-
print(f"Resampling reference audio from {sr} Hz to {target_sample_rate} Hz.")
|
547 |
-
resampler = torchaudio.transforms.Resample(sr, target_sample_rate).to(device)
|
548 |
-
audio = resampler(audio)
|
549 |
-
|
550 |
-
ref_audio_len_frames = audio.shape[-1] // hop_length
|
551 |
-
print(f"Reference audio length: {audio.shape[-1]/target_sample_rate:.2f}s ({ref_audio_len_frames} frames)")
|
552 |
-
|
553 |
-
generated_waves = []
|
554 |
-
spectrograms = []
|
555 |
-
|
556 |
-
progress_bar = tqdm(gen_text_ipa_batches, desc="Generating Batches")
|
557 |
-
for i, gen_text_ipa in enumerate(progress_bar):
|
558 |
-
progress_bar.set_postfix({"Batch": f"{i+1}/{len(gen_text_ipa_batches)}"})
|
559 |
-
|
560 |
-
# Combine reference and generated IPA text
|
561 |
-
combined_ipa_text = ref_text_ipa + " " + gen_text_ipa
|
562 |
-
# print(f"Batch {i+1} Combined IPA: {combined_ipa_text}") # Debug
|
563 |
-
|
564 |
-
# Tokenize
|
565 |
-
try:
|
566 |
-
# Tokenizer expects single string or list of strings
|
567 |
-
encoding = tokenizer.encode(combined_ipa_text)
|
568 |
-
tokens = encoding.ids
|
569 |
-
token_str = encoding.tokens # For logging/debug
|
570 |
-
|
571 |
-
# --- Model Input Formatting ---
|
572 |
-
# Check how your specific model's `sample` method expects the 'text' input.
|
573 |
-
# Option 1 (like app.py): String of space-separated tokens
|
574 |
-
# token_input_string = ' '.join(map(str, token_str))
|
575 |
-
# final_text_list = [token_input_string]
|
576 |
-
|
577 |
-
# Option 2: List of token IDs (might be more common)
|
578 |
-
# final_text_list = [tokens] # List containing the list/tensor of IDs
|
579 |
-
|
580 |
-
# Option 3: Tensor of token IDs (check model docs)
|
581 |
-
# Assuming model expects Option 1 based on app.py:
|
582 |
-
token_input_string = ' '.join(map(str, token_str))
|
583 |
-
final_text_list = [token_input_string]
|
584 |
-
# print(f"Batch {i+1} Input Text List for Model: {final_text_list}")
|
585 |
-
|
586 |
-
except Exception as e:
|
587 |
-
print(f"Error tokenizing batch {i+1}: '{combined_ipa_text}'. Error: {e}")
|
588 |
-
continue
|
589 |
-
|
590 |
-
# Calculate duration
|
591 |
-
ref_ipa_len = len(ref_text_ipa)
|
592 |
-
gen_ipa_len = len(gen_text_ipa)
|
593 |
-
if ref_ipa_len == 0: ref_ipa_len = 1 # Avoid division by zero
|
594 |
-
|
595 |
-
duration_frames = ref_audio_len_frames + int(((ref_audio_len_frames / ref_ipa_len) * gen_ipa_len) / speed)
|
596 |
-
min_duration_frames = max(10, target_sample_rate // hop_length // 4) # Shorter min duration (e.g. 0.25s)
|
597 |
-
duration_frames = max(min_duration_frames, duration_frames)
|
598 |
-
max_duration_frames = 40 * target_sample_rate // hop_length # Increase max duration slightly?
|
599 |
-
if duration_frames > max_duration_frames:
|
600 |
-
print(f"Warning: Calculated duration {duration_frames} frames exceeds max {max_duration_frames}. Capping.")
|
601 |
-
duration_frames = max_duration_frames
|
602 |
-
|
603 |
-
# print(f"Batch {i+1}: Duration={duration_frames} frames")
|
604 |
-
|
605 |
-
# Inference
|
606 |
-
try:
|
607 |
-
with torch.inference_mode():
|
608 |
-
cond_audio = audio.to(ema_model.device, dtype=dtype) # Match model device/dtype
|
609 |
-
# print(f"Model device: {ema_model.device}, Cond audio device: {cond_audio.device}, dtype: {cond_audio.dtype}")
|
610 |
-
|
611 |
-
generated_mel, _ = ema_model.sample(
|
612 |
-
cond=cond_audio,
|
613 |
-
text=final_text_list, # Pass formatted text input
|
614 |
-
duration=duration_frames,
|
615 |
-
steps=nfe_step,
|
616 |
-
cfg_strength=cfg_strength,
|
617 |
-
sway_sampling_coef=sway_sampling_coef,
|
618 |
-
)
|
619 |
-
|
620 |
-
# Process generated mel
|
621 |
-
generated_mel = generated_mel.to(device, dtype=dtype) # Back to main device/dtype
|
622 |
-
generated_mel = generated_mel[:, ref_audio_len_frames:, :]
|
623 |
-
generated_mel_spec = rearrange(generated_mel, "1 n d -> 1 d n")
|
624 |
-
|
625 |
-
# Vocoding
|
626 |
-
# Vocos usually expects float32
|
627 |
-
vocos_input_mel = generated_mel_spec.to(vocos.device, dtype=torch.float32)
|
628 |
-
generated_wave = vocos.decode(vocos_input_mel)
|
629 |
-
generated_wave = generated_wave.to(device, dtype=torch.float32)
|
630 |
-
|
631 |
-
# Adjust RMS (Scale generated audio by the same factor applied to reference)
|
632 |
-
generated_wave = generated_wave * rms_applied_factor
|
633 |
-
|
634 |
-
# Convert to numpy
|
635 |
-
generated_wave_np = generated_wave.squeeze().cpu().numpy()
|
636 |
-
generated_waves.append(generated_wave_np)
|
637 |
-
spectrograms.append(generated_mel_spec[0].cpu().to(torch.float32).numpy())
|
638 |
-
|
639 |
-
except Exception as e:
|
640 |
-
logging.exception(f"Error during inference/processing for batch {i+1}:") # Log traceback
|
641 |
-
print(f"Error details: {e}")
|
642 |
-
continue
|
643 |
-
|
644 |
-
if not generated_waves:
|
645 |
-
print("No audio waves were generated.")
|
646 |
-
return None, None
|
647 |
-
|
648 |
-
# Combine batches
|
649 |
-
print(f"Combining {len(generated_waves)} generated batches...")
|
650 |
-
if cross_fade_duration <= 0 or len(generated_waves) == 1:
|
651 |
-
final_wave = np.concatenate(generated_waves)
|
652 |
-
else:
|
653 |
-
# (Cross-fading logic remains the same)
|
654 |
-
final_wave = generated_waves[0]
|
655 |
-
for i in range(1, len(generated_waves)):
|
656 |
-
prev_wave = final_wave; next_wave = generated_waves[i]
|
657 |
-
cf_samples = min(int(cross_fade_duration * target_sample_rate), len(prev_wave), len(next_wave))
|
658 |
-
if cf_samples <= 0: final_wave = np.concatenate([prev_wave, next_wave]); continue
|
659 |
-
p_olap = prev_wave[-cf_samples:]; n_olap = next_wave[:cf_samples]
|
660 |
-
f_out = np.linspace(1, 0, cf_samples, dtype=p_olap.dtype); f_in = np.linspace(0, 1, cf_samples, dtype=n_olap.dtype)
|
661 |
-
cf_olap = p_olap * f_out + n_olap * f_in
|
662 |
-
final_wave = np.concatenate([prev_wave[:-cf_samples], cf_olap, next_wave[cf_samples:]])
|
663 |
-
print(f"Applied cross-fade of {cross_fade_duration:.2f}s between batches.")
|
664 |
-
|
665 |
-
# Optional: Remove silence post-combination
|
666 |
-
if remove_silence_post:
|
667 |
-
print("Removing silence from final output...")
|
668 |
-
try:
|
669 |
-
final_wave_float32 = final_wave.astype(np.float32)
|
670 |
-
with tempfile.NamedTemporaryFile(delete=True, suffix=".wav") as tmp_wav:
|
671 |
-
sf.write(tmp_wav.name, final_wave_float32, target_sample_rate)
|
672 |
-
aseg = AudioSegment.from_file(tmp_wav.name)
|
673 |
-
non_silent_segs = silence.split_on_silence(
|
674 |
-
aseg, min_silence_len=1000, silence_thresh=-50, keep_silence=500
|
675 |
-
)
|
676 |
-
if not non_silent_segs:
|
677 |
-
print("Warning: Silence removal resulted in empty audio. Keeping original.")
|
678 |
-
else:
|
679 |
-
non_silent_wave = sum(non_silent_segs, AudioSegment.silent(duration=0))
|
680 |
-
non_silent_wave.export(tmp_wav.name, format="wav")
|
681 |
-
final_wave_tensor, _ = torchaudio.load(tmp_wav.name)
|
682 |
-
final_wave = final_wave_tensor.squeeze().cpu().numpy()
|
683 |
-
print("Silence removal applied.")
|
684 |
-
except Exception as e:
|
685 |
-
print(f"Warning: Failed to remove silence: {e}. Using original.")
|
686 |
-
|
687 |
-
# Combine spectrograms
|
688 |
-
print("Combining spectrograms...")
|
689 |
-
try:
|
690 |
-
if spectrograms:
|
691 |
-
combined_spectrogram = np.concatenate(spectrograms, axis=1)
|
692 |
-
else:
|
693 |
-
combined_spectrogram = None
|
694 |
-
except ValueError as e:
|
695 |
-
print(f"Warning: Could not concatenate spectrograms: {e}. Skipping.")
|
696 |
-
combined_spectrogram = None
|
697 |
-
|
698 |
-
return final_wave, combined_spectrogram
|
699 |
-
|
700 |
-
|
701 |
-
def main_infer(ref_audio_orig_path, ref_text_input, gen_text_full,
|
702 |
-
ema_model, vocos, tokenizer, pipe_asr, # Loaded models/utils
|
703 |
-
ref_language, language, # Languages
|
704 |
-
speed, nfe_step, cfg_strength, sway_sampling_coef, # Sampling params
|
705 |
-
remove_silence_flag, cross_fade_duration, # Postprocessing
|
706 |
-
target_sample_rate, hop_length, target_rms, # Audio params
|
707 |
-
device, dtype): # System params
|
708 |
-
"""
|
709 |
-
Main inference function coordinating preprocessing, batching, and generation.
|
710 |
-
(Function body remains the same as previous refactored version)
|
711 |
-
"""
|
712 |
-
print(f"Starting inference for text: '{gen_text_full[:100]}...'")
|
713 |
-
|
714 |
-
# --- Reference Audio Preprocessing ---
|
715 |
-
print("Processing reference audio...")
|
716 |
-
processed_ref_path = None
|
717 |
-
try:
|
718 |
-
with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as temp_ref_wav:
|
719 |
-
processed_ref_path = temp_ref_wav.name # Store path for potential use
|
720 |
-
aseg = AudioSegment.from_file(ref_audio_orig_path)
|
721 |
-
print(f"Original ref duration: {len(aseg)/1000:.2f}s")
|
722 |
-
|
723 |
-
# Edge silence removal + padding
|
724 |
-
aseg = remove_silence_edges(aseg)
|
725 |
-
aseg += AudioSegment.silent(duration=150)
|
726 |
-
|
727 |
-
# Split/recombine on silence
|
728 |
-
non_silent_segs = silence.split_on_silence(
|
729 |
-
aseg, min_silence_len=700, silence_thresh=-50, keep_silence=700
|
730 |
-
)
|
731 |
-
if non_silent_segs:
|
732 |
-
aseg = sum(non_silent_segs, AudioSegment.silent(duration=0)) # Use sum for conciseness
|
733 |
-
else:
|
734 |
-
print("Warning: Silence splitting/recombining resulted in empty audio. Using edge-trimmed.")
|
735 |
-
|
736 |
-
# Clip to 10s
|
737 |
-
max_ref_duration_ms = 10000
|
738 |
-
if len(aseg) > max_ref_duration_ms:
|
739 |
-
print(f"Reference audio exceeds {max_ref_duration_ms/1000}s. Clipping...")
|
740 |
-
aseg = aseg[:max_ref_duration_ms]
|
741 |
-
|
742 |
-
aseg.export(processed_ref_path, format="wav")
|
743 |
-
print(f"Processed ref duration: {len(aseg)/1000:.2f}s. Saved to temp file: {processed_ref_path}")
|
744 |
-
|
745 |
-
# Load processed audio tensor
|
746 |
-
ref_audio_tensor, sr_ref = torchaudio.load(processed_ref_path)
|
747 |
-
|
748 |
-
except Exception as e:
|
749 |
-
print(f"Error processing reference audio {ref_audio_orig_path}: {e}")
|
750 |
-
if processed_ref_path and Path(processed_ref_path).exists():
|
751 |
-
Path(processed_ref_path).unlink() # Clean up temp file on error
|
752 |
-
raise
|
753 |
-
|
754 |
-
# --- Reference Text Handling ---
|
755 |
-
ref_text_processed = ""
|
756 |
-
if not ref_text_input or ref_text_input.strip() == "":
|
757 |
-
print("No reference text provided. Transcribing reference audio...")
|
758 |
-
if pipe_asr is None:
|
759 |
-
raise ValueError("Whisper ASR model not loaded. Cannot transcribe. Please provide --ref_text.")
|
760 |
-
if not processed_ref_path:
|
761 |
-
raise ValueError("Processed reference audio path is missing for transcription.")
|
762 |
-
try:
|
763 |
-
# Ensure Whisper input dtype matches its loaded dtype
|
764 |
-
whisper_input_dtype = pipe_asr.model.dtype
|
765 |
-
|
766 |
-
# Load audio specifically for Whisper if dtypes differ significantly
|
767 |
-
# Or rely on pipeline handling. Assuming pipeline handles it for now.
|
768 |
-
print(f"Transcribing: {processed_ref_path}")
|
769 |
-
transcription_result = pipe_asr(
|
770 |
-
processed_ref_path,
|
771 |
-
chunk_length_s=15,
|
772 |
-
batch_size=8, # Smaller batch size for CLI
|
773 |
-
generate_kwargs={"task": "transcribe", "language": None}, # Whisper language detection
|
774 |
-
return_timestamps=False,
|
775 |
-
)
|
776 |
-
ref_text_processed = transcription_result["text"].strip()
|
777 |
-
print(f"Transcription finished: '{ref_text_processed}'")
|
778 |
-
if not ref_text_processed:
|
779 |
-
print("Warning: Transcription resulted in empty text. Using placeholder.")
|
780 |
-
ref_text_processed = "Reference audio"
|
781 |
-
except Exception as e:
|
782 |
-
logging.exception("Error during transcription:")
|
783 |
-
raise ValueError("Transcription failed. Please provide --ref_text.")
|
784 |
-
else:
|
785 |
-
print("Using provided reference text.")
|
786 |
-
ref_text_processed = ref_text_input
|
787 |
-
|
788 |
-
# Clean up the temporary processed reference audio file
|
789 |
-
if processed_ref_path and Path(processed_ref_path).exists():
|
790 |
-
try:
|
791 |
-
Path(processed_ref_path).unlink()
|
792 |
-
# print(f"Cleaned up temp ref file: {processed_ref_path}") # Debug
|
793 |
-
except OSError as e:
|
794 |
-
print(f"Warning: Could not delete temp ref file {processed_ref_path}: {e}")
|
795 |
-
|
796 |
-
|
797 |
-
# Ensure reference text ends with ". "
|
798 |
-
if not ref_text_processed.endswith(". "):
|
799 |
-
ref_text_processed = ref_text_processed.rstrip('. ') + ". " # More robust way
|
800 |
-
print(f"Final Reference Text: '{ref_text_processed}'")
|
801 |
-
|
802 |
-
# --- Phonemize Reference Text ---
|
803 |
-
print(f"Phonemizing reference text with language: {ref_language}")
|
804 |
-
ref_text_ipa = text_to_ipa(ref_text_processed, language=ref_language)
|
805 |
-
if not ref_text_ipa: raise ValueError("Reference text phonemization failed.")
|
806 |
-
|
807 |
-
# --- Chunk and Phonemize Generation Text ---
|
808 |
-
ref_audio_duration_sec = ref_audio_tensor.shape[-1] / sr_ref if sr_ref > 0 else 1.0
|
809 |
-
if ref_audio_duration_sec <= 0: ref_audio_duration_sec = 1.0
|
810 |
-
chars_per_sec = len(ref_text_processed.encode('utf-8')) / ref_audio_duration_sec if ref_audio_duration_sec > 0 else 10.0
|
811 |
-
if chars_per_sec <= 0: chars_per_sec = 10.0
|
812 |
-
target_chunk_duration_sec = max(5.0, 20.0 - ref_audio_duration_sec)
|
813 |
-
max_chars = int(chars_per_sec * target_chunk_duration_sec)
|
814 |
-
|
815 |
-
print(f"Ref duration: {ref_audio_duration_sec:.2f}s => Calculated max_chars/batch: {max_chars}")
|
816 |
-
gen_text_batches_plain = chunk_text(gen_text_full, max_chars=max_chars)
|
817 |
-
if not gen_text_batches_plain: raise ValueError("Text chunking resulted in zero batches.")
|
818 |
-
print(f"Split generation text into {len(gen_text_batches_plain)} batches.")
|
819 |
-
|
820 |
-
print(f"Phonemizing generation text batches with language: {language}")
|
821 |
-
gen_text_ipa_batches = []
|
822 |
-
for i, batch_text in enumerate(gen_text_batches_plain):
|
823 |
-
# print(f" Phonemizing batch {i+1}/{len(gen_text_batches_plain)}...") # Verbose
|
824 |
-
batch_ipa = text_to_ipa(batch_text, language=language)
|
825 |
-
if batch_ipa: gen_text_ipa_batches.append(batch_ipa)
|
826 |
-
else: print(f"Warning: Skipping batch {i+1} due to phonemization failure.")
|
827 |
-
|
828 |
-
if not gen_text_ipa_batches: raise ValueError("Phonemization failed for all generation text batches.")
|
829 |
-
|
830 |
-
# --- Run Batched Inference ---
|
831 |
-
print(f"Starting batch inference process ({len(gen_text_ipa_batches)} batches)...")
|
832 |
-
final_wave, combined_spectrogram = infer_batch(
|
833 |
-
(ref_audio_tensor, sr_ref), ref_text_ipa, gen_text_ipa_batches,
|
834 |
-
ema_model, vocos, tokenizer,
|
835 |
-
remove_silence_flag, cross_fade_duration,
|
836 |
-
nfe_step, cfg_strength, sway_sampling_coef, speed,
|
837 |
-
target_sample_rate, hop_length, target_rms,
|
838 |
-
device, dtype
|
839 |
-
)
|
840 |
-
|
841 |
-
return final_wave, combined_spectrogram
|
842 |
-
|
843 |
-
|
844 |
-
# --- Execution ---
|
845 |
if __name__ == "__main__":
|
846 |
-
|
847 |
-
logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s')
|
848 |
-
|
849 |
-
try:
|
850 |
-
final_wave_np, combined_spectrogram_np = main_infer(
|
851 |
-
ref_audio_path, ref_text, gen_text,
|
852 |
-
ema_model, vocos, tokenizer, pipe,
|
853 |
-
ref_language, language,
|
854 |
-
speed, nfe_step, cfg_strength, sway_sampling_coef,
|
855 |
-
remove_silence_flag, cross_fade_duration,
|
856 |
-
target_sample_rate, hop_length, target_rms,
|
857 |
-
device, dtype
|
858 |
-
)
|
859 |
-
|
860 |
-
# --- Save Outputs ---
|
861 |
-
output_saved = False
|
862 |
-
if final_wave_np is not None and len(final_wave_np) > 0:
|
863 |
-
print(f"Saving final audio ({len(final_wave_np)/target_sample_rate:.2f}s) to {wave_path}...")
|
864 |
-
final_wave_float32 = final_wave_np.astype(np.float32) # Ensure float32 for sf
|
865 |
-
sf.write(str(wave_path), final_wave_float32, target_sample_rate)
|
866 |
-
print("Audio saved successfully.")
|
867 |
-
output_saved = True
|
868 |
-
else:
|
869 |
-
print("Inference did not produce a valid audio wave.")
|
870 |
-
|
871 |
-
if combined_spectrogram_np is not None:
|
872 |
-
print(f"Saving combined spectrogram to {spectrogram_path}...")
|
873 |
-
save_spectrogram(combined_spectrogram_np, str(spectrogram_path))
|
874 |
-
print("Spectrogram saved successfully.")
|
875 |
-
output_saved = True
|
876 |
-
# else: # No need to print if spectrogram was None
|
877 |
-
# print("Spectrogram generation failed or was skipped.")
|
878 |
-
|
879 |
-
if not output_saved:
|
880 |
-
print("No output files were generated.")
|
881 |
-
|
882 |
-
except FileNotFoundError as e:
|
883 |
-
logging.error(f"File not found: {e}")
|
884 |
-
print(f"\nError: A required file was not found. Please check paths. Details: {e}")
|
885 |
-
exit(1)
|
886 |
-
except ValueError as e:
|
887 |
-
logging.error(f"Value error: {e}")
|
888 |
-
print(f"\nError: An invalid value or configuration was encountered. Details: {e}")
|
889 |
-
exit(1)
|
890 |
-
except Exception as e:
|
891 |
-
logging.exception("An unexpected error occurred during inference:") # Log traceback
|
892 |
-
print(f"\nAn unexpected error occurred: {e}")
|
893 |
-
exit(1)
|
894 |
|
895 |
-
|
|
|
1 |
+
# --- START OF FILE inference_cli.py ---
|
2 |
+
|
3 |
import argparse
|
4 |
+
import shutil
|
|
|
|
|
|
|
|
|
|
|
5 |
import soundfile as sf
|
6 |
+
import os # For path manipulation if needed
|
7 |
+
import sys # To potentially add app.py directory to path
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
8 |
|
9 |
+
# Try to import app.py - assumes it's in the same directory or Python path
|
|
|
10 |
try:
|
11 |
+
# If app.py is not directly importable, you might need to add its directory to the path
|
12 |
+
# Example: sys.path.append(os.path.dirname(os.path.abspath(__file__))) # Add current dir
|
13 |
+
import app
|
14 |
+
from app import infer # Import the main inference function
|
15 |
+
except ImportError as e:
|
16 |
+
print(f"Error: Could not import 'app.py'. Make sure it's in the Python path.")
|
17 |
+
print(f"Details: {e}")
|
18 |
+
sys.exit(1)
|
19 |
+
except Exception as e:
|
20 |
+
print(f"An unexpected error occurred during 'app.py' import: {e}")
|
21 |
+
sys.exit(1)
|
22 |
+
|
23 |
+
|
24 |
+
def main():
|
25 |
+
parser = argparse.ArgumentParser(description="F5 TTS - Simplified CLI Interface using app.py")
|
26 |
+
|
27 |
+
# --- Input Arguments ---
|
28 |
+
parser.add_argument("--ref_audio", required=True, help="Path to the reference audio file (wav, mp3, etc.)")
|
29 |
+
parser.add_argument("--ref_text", default="", help="Reference text. If empty, audio transcription will be performed by app.py's infer function.")
|
30 |
+
parser.add_argument("--gen_text", required=True, help="Text to generate")
|
31 |
+
|
32 |
+
# --- Model & Generation Parameters ---
|
33 |
+
# Note: app.py seems hardcoded to load the "Multi" model at the top level.
|
34 |
+
# This argument might not change the loaded model unless app.py's infer logic uses it internally.
|
35 |
+
parser.add_argument("--exp_name", default="Multi", help="Experiment name / model selection (default: Multi - effectiveness depends on app.py)")
|
36 |
+
parser.add_argument("--language", default="en-us", help="Synthesized language code (e.g., en-us, pl, de) (default: en-us)")
|
37 |
+
parser.add_argument("--ref_language", default="en-us", help="Reference language code (e.g., en-us, pl, de) (default: en-us)")
|
38 |
+
parser.add_argument("--speed", type=float, default=1.0, help="Audio speed factor (default: 1.0)")
|
39 |
+
|
40 |
+
# --- Postprocessing ---
|
41 |
+
parser.add_argument("--remove_silence", action="store_true", help="Remove silence from the output audio (uses app.py logic)")
|
42 |
+
parser.add_argument("--cross_fade_duration", type=float, default=0.15, help="Cross-fade duration between batches (s)")
|
43 |
+
|
44 |
+
# --- Output Arguments ---
|
45 |
+
parser.add_argument("--output_audio", default="output.wav", help="Path to save the output WAV file")
|
46 |
+
parser.add_argument("--output_spectrogram", default="spectrogram.png", help="Path to save the spectrogram image (PNG)")
|
47 |
+
|
48 |
+
args = parser.parse_args()
|
49 |
+
|
50 |
+
print("--- Configuration ---")
|
51 |
+
print(f"Reference Audio: {args.ref_audio}")
|
52 |
+
print(f"Reference Text: '{args.ref_text if args.ref_text else '<Automatic Transcription>'}'")
|
53 |
+
print(f"Generation Text: '{args.gen_text[:100]}...'")
|
54 |
+
print(f"Model (exp_name): {args.exp_name}")
|
55 |
+
print(f"Synth Language: {args.language}")
|
56 |
+
print(f"Ref Language: {args.ref_language}")
|
57 |
+
print(f"Speed: {args.speed}")
|
58 |
+
print(f"Remove Silence: {args.remove_silence}")
|
59 |
+
print(f"Cross-Fade: {args.cross_fade_duration}s")
|
60 |
+
print(f"Output Audio: {args.output_audio}")
|
61 |
+
print(f"Output Spectrogram: {args.output_spectrogram}")
|
62 |
+
print("--------------------")
|
63 |
+
|
64 |
+
# --- Set Global Variables in app.py ---
|
65 |
+
# The 'infer' function in app.py relies on these globals being set.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
66 |
try:
|
67 |
+
print(f"Setting language in app module to: {args.language}")
|
68 |
+
app.language = args.language
|
69 |
+
print(f"Setting ref_language in app module to: {args.ref_language}")
|
70 |
+
app.ref_language = args.ref_language
|
71 |
+
print(f"Setting speed in app module to: {args.speed}")
|
72 |
+
app.speed = args.speed
|
73 |
+
except AttributeError as e:
|
74 |
+
print(f"Error: Could not set global variable in 'app.py'. Does it exist? Details: {e}")
|
75 |
+
sys.exit(1)
|
76 |
+
|
77 |
+
# --- Run Inference ---
|
78 |
+
print("\nStarting inference process (will load models if not already loaded)...")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
79 |
try:
|
80 |
+
# Call the infer function directly from the imported app module
|
81 |
+
(sr, audio_data), temp_spectrogram_path = infer(
|
82 |
+
ref_audio_orig=args.ref_audio,
|
83 |
+
ref_text=args.ref_text,
|
84 |
+
gen_text=args.gen_text,
|
85 |
+
exp_name=args.exp_name,
|
86 |
+
remove_silence=args.remove_silence,
|
87 |
+
cross_fade_duration=args.cross_fade_duration
|
88 |
+
# Note: language, ref_language, speed are used globally within app.py's functions
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
89 |
)
|
90 |
+
print("Inference completed.")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
91 |
|
|
|
|
|
|
|
|
|
|
|
92 |
except Exception as e:
|
93 |
+
print(f"\nError during inference: {e}")
|
94 |
+
import traceback
|
95 |
+
traceback.print_exc() # Print detailed traceback
|
96 |
+
sys.exit(1)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
97 |
|
98 |
+
# --- Save Outputs ---
|
|
|
|
|
99 |
try:
|
100 |
+
# Save audio
|
101 |
+
print(f"Saving audio to: {args.output_audio}")
|
102 |
+
# Ensure directory exists
|
103 |
+
os.makedirs(os.path.dirname(os.path.abspath(args.output_audio)) or '.', exist_ok=True)
|
104 |
+
# Ensure data is float32 for soundfile
|
105 |
+
if audio_data.dtype != "float32":
|
106 |
+
audio_data = audio_data.astype("float32")
|
107 |
+
sf.write(args.output_audio, audio_data, sr)
|
108 |
+
|
109 |
+
# Copy spectrogram from the temporary path returned by infer
|
110 |
+
print(f"Copying spectrogram from {temp_spectrogram_path} to: {args.output_spectrogram}")
|
111 |
+
# Ensure directory exists
|
112 |
+
os.makedirs(os.path.dirname(os.path.abspath(args.output_spectrogram)) or '.', exist_ok=True)
|
113 |
+
shutil.copy(temp_spectrogram_path, args.output_spectrogram)
|
114 |
+
|
115 |
+
print("\n--- Success ---")
|
116 |
+
print(f"Audio saved in: {args.output_audio}")
|
117 |
+
print(f"Spectrogram saved in: {args.output_spectrogram}")
|
118 |
+
print("---------------")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
119 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
120 |
except Exception as e:
|
121 |
+
print(f"\nError saving output files: {e}")
|
122 |
+
sys.exit(1)
|
|
|
|
|
|
|
|
|
|
|
123 |
|
124 |
+
# Optional: Clean up the temporary spectrogram file if needed,
|
125 |
+
# but NamedTemporaryFile usually handles this if delete=True was used in app.py
|
126 |
+
# try:
|
127 |
+
# if os.path.exists(temp_spectrogram_path):
|
128 |
+
# os.remove(temp_spectrogram_path)
|
129 |
+
# except Exception as e:
|
130 |
+
# print(f"Warning: Could not clean up temporary spectrogram file {temp_spectrogram_path}: {e}")
|
|
|
|
|
|
|
131 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
132 |
if __name__ == "__main__":
|
133 |
+
main()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
134 |
|
135 |
+
# --- END OF FILE inference_cli.py ---
|