File size: 4,435 Bytes
0ce1e8d |
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 |
"""
Demo to run OpenAI Whisper using HuggingFace ZeroGPU.
This way we can test default Whisper models provided by OpenAI, for later comparison with fine-tuned ones.
"""
import subprocess
import tempfile
from pathlib import Path
import gradio as gr
import spaces
import torch
import whisper
YT_AUDIO_FORMAT = "bestaudio[ext=m4a]"
def download_youtube(url: str, tmp_dir: Path) -> Path:
"""Download the audio track from a YouTube video and return the local path."""
out_path = tmp_dir / r"%\(id)s.%(ext)s"
cmd = [
"yt-dlp",
"--quiet",
"--no-warnings",
"--extract-audio",
"--audio-format",
"m4a",
"--audio-quality",
"0",
"-f",
YT_AUDIO_FORMAT,
"-o",
str(out_path),
url,
]
result = subprocess.run(cmd, capture_output=True, check=True)
if result.returncode != 0:
raise RuntimeError(f"yt-dlp failed: {result.stderr.decode()}")
files = list(tmp_dir.glob("*.m4a"))
if not files:
raise FileNotFoundError("Could not locate downloaded audio.")
return files[0]
def _get_input_path(audio, youtube_url):
if youtube_url and youtube_url.strip():
with tempfile.TemporaryDirectory() as tmp:
return download_youtube(youtube_url, Path(tmp))
elif audio is not None:
return audio
else:
raise gr.Error("Provide audio or a YouTube URL")
def make_results_table(results):
rows = []
for r in results:
row = [r["model"], r["language"], r["text"]]
rows.append(row)
return rows
@spaces.GPU
def transcribe_audio(
model_sizes: list[str],
audio: str,
youtube_url: str,
return_timestamps: bool,
temperature: float,
):
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
results = []
for size in model_sizes:
model = whisper.load_model(size, device=device)
inp = _get_input_path(audio, youtube_url)
out = model.transcribe(
str(inp),
word_timestamps=return_timestamps,
temperature=temperature,
verbose=False,
)
text = out["text"].strip()
segments = out["segments"] if return_timestamps else []
results.append(
{
"model": size,
"language": out["language"],
"text": text,
"segments": segments,
}
)
df_results = make_results_table(results, return_timestamps)
return df_results
def build_demo() -> gr.Blocks:
with gr.Blocks(title="🗣️ Whisper Transcription Demo (HF Spaces Zero-GPU)") as whisper_demo:
gr.Markdown("""
# Whisper Transcription Demo
Run Whisper transcription on audio or YouTube video. Whisper is a general-purpose speech recognition model,
trained on a large dataset
""")
with gr.Row():
model_choices = gr.Dropdown(
label="Model size(s)",
choices=["tiny", "base", "small", "medium", "large", "turbo"],
value=["turbo"],
multiselect=True,
allow_custom_value=False,
)
ts_checkbox = gr.Checkbox(
label="Return word timestamps",
interactive=False,
value=False,
)
temp_slider = gr.Slider(
label="Decoding temperature",
minimum=0.0,
maximum=1.0,
value=0.0,
step=0.01,
)
audio_input = gr.Audio(
label="Upload or record audio",
sources=["upload"],
type="filepath",
)
yt_input = gr.Textbox(
label="... or paste a YouTube URL (audio only)",
placeholder="https://youtu.be/XYZ",
)
with gr.Row():
transcribe_btn = gr.Button("Transcribe 🏁")
out_table = gr.Dataframe(
headers=["Model", "Language", "Transcript"],
datatype=["str", "str", "str"],
label="Transcription Results",
)
transcribe_btn.click(
transcribe_audio,
inputs=[model_choices, audio_input, yt_input, ts_checkbox, temp_slider],
outputs=[out_table],
)
return whisper_demo
if __name__ == "__main__":
demo = build_demo()
demo.launch()
|