File size: 9,279 Bytes
6b90e11 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 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 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 | import argparse
from typing import Optional
import datasets
from datasets import Audio
import evaluate
import soundfile as sf
import tempfile
import time
import os
import requests
import itertools
from tqdm import tqdm
from dotenv import load_dotenv
from normalizer import data_utils
from normalizer.eval_utils import normalize_compound_pairs
import concurrent.futures
from providers import get_provider, PermanentError
load_dotenv()
def fetch_audio_urls(dataset_path, config_name, split, batch_size=100, max_retries=20):
API_URL = "https://datasets-server.huggingface.co/rows"
size_url = f"https://datasets-server.huggingface.co/size?dataset={dataset_path}&config={config_name}&split={split}"
size_response = requests.get(size_url).json()
total_rows = size_response["size"]["config"]["num_rows"]
for offset in tqdm(range(0, total_rows, batch_size), desc="Fetching audio URLs"):
params = {
"dataset": dataset_path,
"config": config_name,
"split": split,
"offset": offset,
"length": min(batch_size, total_rows - offset),
}
retries = 0
while retries <= max_retries:
try:
headers = {}
if os.environ.get("HF_TOKEN") is not None:
headers["Authorization"] = f"Bearer {os.environ['HF_TOKEN']}"
else:
print("HF_TOKEN not set, might experience rate-limiting.")
response = requests.get(API_URL, params=params)
response.raise_for_status()
data = response.json()
yield from data["rows"]
break
except (requests.exceptions.RequestException, ValueError) as e:
retries += 1
print(
f"Error fetching data: {e}, retrying ({retries}/{max_retries})..."
)
time.sleep(10)
if retries >= max_retries:
raise Exception("Max retries exceeded while fetching data.")
def transcribe_with_retry(
model_name: str,
audio_file_path: Optional[str],
sample: dict,
max_retries=10,
use_url=False,
language="en",
):
provider, variant = get_provider(model_name)
retries = 0
while retries <= max_retries:
try:
return provider.transcribe(variant, audio_file_path, sample, use_url=use_url, language=language)
except PermanentError:
raise
except Exception as e:
retries += 1
if retries > max_retries:
raise e
if not use_url:
sf.write(
audio_file_path,
sample["audio"]["array"],
sample["audio"]["sampling_rate"],
format="WAV",
)
delay = 1
print(
f"API Error: {str(e)}. Retrying in {delay}s... (Attempt {retries}/{max_retries})"
)
time.sleep(delay)
def transcribe_dataset(
dataset_path,
config_name,
split,
model_name,
language,
use_url=False,
max_samples=None,
max_workers=4,
):
if use_url:
audio_rows = fetch_audio_urls(dataset_path, config_name, split)
if max_samples:
audio_rows = itertools.islice(audio_rows, max_samples)
ds = audio_rows
else:
ds = datasets.load_dataset(dataset_path, config_name, split=split, streaming=False)
ds = ds.cast_column("audio", Audio(sampling_rate=16000))
if max_samples:
ds = ds.select(range(min(max_samples, len(ds))))
results = {
"references": [],
"predictions": [],
"audio_length_s": [],
"transcription_time_s": [],
}
print(f"Transcribing with model: {model_name}, language: {language}, config: {config_name}")
def process_sample(sample):
if use_url:
reference = sample["row"]["text"].strip()
audio_duration = sample["row"]["audio_length_s"]
start = time.time()
try:
transcription = transcribe_with_retry(
model_name, None, sample, use_url=True, language=language
)
except Exception as e:
print(f"Failed to transcribe after retries: {e}")
return None
else:
reference = sample.get("text", "").strip()
with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as tmpfile:
sf.write(
tmpfile.name,
sample["audio"]["array"],
sample["audio"]["sampling_rate"],
format="WAV",
)
tmp_path = tmpfile.name
audio_duration = (
len(sample["audio"]["array"]) / sample["audio"]["sampling_rate"]
)
start = time.time()
try:
transcription = transcribe_with_retry(
model_name, tmp_path, sample, use_url=False, language=language
)
except Exception as e:
print(f"Failed to transcribe after retries: {e}")
os.unlink(tmp_path)
return None
finally:
if os.path.exists(tmp_path):
os.unlink(tmp_path)
else:
print(f"File {tmp_path} does not exist")
transcription_time = time.time() - start
return reference, transcription, audio_duration, transcription_time
with concurrent.futures.ThreadPoolExecutor(max_workers=max_workers) as executor:
future_to_sample = {
executor.submit(process_sample, sample): sample for sample in ds
}
for future in tqdm(
concurrent.futures.as_completed(future_to_sample),
total=len(future_to_sample),
desc="Transcribing",
):
result = future.result()
if result:
reference, transcription, audio_duration, transcription_time = result
results["predictions"].append(transcription)
results["references"].append(reference)
results["audio_length_s"].append(audio_duration)
results["transcription_time_s"].append(transcription_time)
# Filter empty references (consistent with English pipeline's prepare_data)
filtered = [
(ref, pred, dur, time_s)
for ref, pred, dur, time_s in zip(
results["references"], results["predictions"],
results["audio_length_s"], results["transcription_time_s"]
)
if data_utils.is_target_text_in_range(ref)
]
if filtered:
results["references"], results["predictions"], results["audio_length_s"], results["transcription_time_s"] = zip(*filtered)
results = {k: list(v) for k, v in results.items()}
manifest_path = data_utils.write_manifest(
results["references"],
results["predictions"],
model_name.replace("/", "-"),
dataset_path,
config_name,
split,
audio_length=results["audio_length_s"],
transcription_time=results["transcription_time_s"],
)
print("Results saved at path:", manifest_path)
norm_refs = [data_utils.ml_normalizer(r, lang=language) for r in results["references"]]
norm_preds = [data_utils.ml_normalizer(t, lang=language) for t in results["predictions"]]
wer_metric = evaluate.load("wer")
wer_refs, wer_preds = normalize_compound_pairs(norm_refs, norm_preds)
wer = wer_metric.compute(references=wer_refs, predictions=wer_preds)
wer_percent = round(100 * wer, 2)
rtfx = round(
sum(results["audio_length_s"]) / sum(results["transcription_time_s"]), 2
)
print("WER:", wer_percent, "%")
print("RTFx:", rtfx)
if __name__ == "__main__":
parser = argparse.ArgumentParser(
description="Multilingual API Transcription Script with Concurrency"
)
parser.add_argument("--dataset_path", required=True)
parser.add_argument("--config_name", required=True, help="Dataset config name, e.g. 'fleurs_de'")
parser.add_argument("--language", required=True, help="Language code, e.g. 'de'")
parser.add_argument("--split", default="test")
parser.add_argument(
"--model_name",
required=True,
help="Prefix model name with provider prefix (e.g., 'assembly/', 'openai/', 'elevenlabs/', 'revai/', 'speechmatics/' or 'aquavoice/')",
)
parser.add_argument("--max_samples", type=int, default=None)
parser.add_argument(
"--max_workers", type=int, default=300, help="Number of concurrent threads"
)
parser.add_argument(
"--use_url",
action="store_true",
help="Use URL-based audio fetching instead of datasets",
)
args = parser.parse_args()
transcribe_dataset(
dataset_path=args.dataset_path,
config_name=args.config_name,
split=args.split,
model_name=args.model_name,
language=args.language,
use_url=args.use_url,
max_samples=args.max_samples,
max_workers=args.max_workers,
)
|