nihongoMiniteto / app.py
machinelearnAn's picture
Update app.py
91c6a66 verified
import os
import gradio as gr
from faster_whisper import WhisperModel
import soundfile as sf
# Fixed config for this app
MODEL_NAME = os.getenv("MODEL_NAME", "Systran/faster-whisper-small") # fast & accurate enough for short clips
LANGUAGE = "ja" # force Japanese
VAD = os.getenv("VAD_FILTER", "1") == "1"
MAX_SECONDS = int(os.getenv("MAX_SECONDS", "120")) # 2 minutes
_model = None
def get_model():
global _model
if _model is not None:
return _model
# GPU first, then CPU fallbacks
for device, compute_type in [("cuda", "float16"), ("cuda", "int8_float16"), ("cpu", "int8")]:
try:
m = WhisperModel(MODEL_NAME, device=device, compute_type=compute_type)
_model = m
print(f"[load] {MODEL_NAME} on {device}/{compute_type}")
return _model
except Exception as e:
print(f"[load-failed] {device}/{compute_type}: {e}")
continue
raise RuntimeError("Unable to load model.")
def transcribe_upload(audio_path):
if not audio_path:
return "ファイルが選択されていません。"
# duration guard
try:
data, sr = sf.read(audio_path)
duration = len(data) / float(sr)
if duration > MAX_SECONDS:
return f"音声が長すぎます({duration:.1f}秒)。最大{MAX_SECONDS}秒のファイルのみ対応しています。"
except Exception as e:
print(f"[warn] duration check failed: {e}")
model = get_model()
segments, info = model.transcribe(
audio_path,
language=LANGUAGE, # 固定: 日本語
task="transcribe",
vad_filter=VAD,
)
text = "".join(seg.text for seg in segments)
return text.strip()
with gr.Blocks() as demo:
gr.Markdown("# 🇯🇵 日本語 音声→テキスト(アップロードのみ)\n- 日本語の音声ファイル(最大2分)をアップロードしてください。\n- 変換後のテキストが下に表示されます。")
audio = gr.Audio(sources=["upload"], type="filepath", label="音声ファイルをアップロード(<2分)")
out = gr.Textbox(lines=8, label="テキスト")
gr.Button("文字起こし").click(transcribe_upload, inputs=[audio], outputs=[out])
demo.launch()