返回 F5-TTS
eval_librispeech_test_clean.py
根目录 / src / f5_tts / eval / eval_librispeech_test_clean.py
1 # Evaluate with Librispeech test-clean, ~3s prompt to generate 4-10s audio (the way of valle/voicebox evaluation)
2
3 import argparse
4 import ast
5 import json
6 import os
7 import sys
8
9
10 sys.path.append(os.getcwd())
11
12 import multiprocessing as mp
13 from importlib.resources import files
14
15 import numpy as np
16
17 from f5_tts.eval.utils_eval import get_librispeech_test, run_asr_wer, run_sim
18
19
20 rel_path = str(files("f5_tts").joinpath("../../"))
21
22
23 def get_args():
24 parser = argparse.ArgumentParser()
25 parser.add_argument("-e", "--eval_task", type=str, default="wer", choices=["sim", "wer"])
26 parser.add_argument("-l", "--lang", type=str, default="en")
27 parser.add_argument("-g", "--gen_wav_dir", type=str, required=True)
28 parser.add_argument("-p", "--librispeech_test_clean_path", type=str, required=True)
29 parser.add_argument(
30 "-n", "--gpu_nums", type=str, default="8", help="Number of GPUs to use (e.g., 8) or GPU list (e.g., [0,1,2,3])"
31 )
32 parser.add_argument("--local", action="store_true", help="Use local custom checkpoint directory")
33 return parser.parse_args()
34
35
36 def parse_gpu_nums(gpu_nums_str):
37 try:
38 if gpu_nums_str.startswith("[") and gpu_nums_str.endswith("]"):
39 gpu_list = ast.literal_eval(gpu_nums_str)
40 if isinstance(gpu_list, list):
41 return gpu_list
42 return list(range(int(gpu_nums_str)))
43 except (ValueError, SyntaxError):
44 raise argparse.ArgumentTypeError(
45 f"Invalid GPU specification: {gpu_nums_str}. Use a number (e.g., 8) or a list (e.g., [0,1,2,3])"
46 )
47
48
49 def main():
50 args = get_args()
51 eval_task = args.eval_task
52 lang = args.lang
53 librispeech_test_clean_path = args.librispeech_test_clean_path # test-clean path
54 gen_wav_dir = args.gen_wav_dir
55 metalst = rel_path + "/data/librispeech_pc_test_clean_cross_sentence.lst"
56
57 gpus = parse_gpu_nums(args.gpu_nums)
58 test_set = get_librispeech_test(metalst, gen_wav_dir, gpus, librispeech_test_clean_path)
59
60 ## In LibriSpeech, some speakers utilized varying voice characteristics for different characters in the book,
61 ## leading to a low similarity for the ground truth in some cases.
62 # test_set = get_librispeech_test(metalst, gen_wav_dir, gpus, librispeech_test_clean_path, eval_ground_truth = True) # eval ground truth
63
64 local = args.local
65 if local: # use local custom checkpoint dir
66 asr_ckpt_dir = "../checkpoints/Systran/faster-whisper-large-v3"
67 else:
68 asr_ckpt_dir = "" # auto download to cache dir
69 wavlm_ckpt_dir = "../checkpoints/UniSpeech/wavlm_large_finetune.pth"
70
71 # --------------------------------------------------------------------------
72
73 full_results = []
74 metrics = []
75
76 if eval_task == "wer":
77 with mp.Pool(processes=len(gpus)) as pool:
78 args = [(rank, lang, sub_test_set, asr_ckpt_dir) for (rank, sub_test_set) in test_set]
79 results = pool.map(run_asr_wer, args)
80 for r in results:
81 full_results.extend(r)
82 elif eval_task == "sim":
83 with mp.Pool(processes=len(gpus)) as pool:
84 args = [(rank, sub_test_set, wavlm_ckpt_dir) for (rank, sub_test_set) in test_set]
85 results = pool.map(run_sim, args)
86 for r in results:
87 full_results.extend(r)
88 else:
89 raise ValueError(f"Unknown metric type: {eval_task}")
90
91 result_path = f"{gen_wav_dir}/_{eval_task}_results.jsonl"
92 with open(result_path, "w") as f:
93 for line in full_results:
94 metrics.append(line[eval_task])
95 f.write(json.dumps(line, ensure_ascii=False) + "\n")
96 metric = round(np.mean(metrics), 5)
97 f.write(f"\n{eval_task.upper()}: {metric}\n")
98
99 print(f"\nTotal {len(metrics)} samples")
100 print(f"{eval_task.upper()}: {metric}")
101 print(f"{eval_task.upper()} results saved to {result_path}")
102
103
104 if __name__ == "__main__":
105 main()
106
106 lines PYTHON