Upload 8 files
Browse files- app.py +14 -13
- tests/test_zerogpu_config.py +8 -0
app.py
CHANGED
|
@@ -30,9 +30,9 @@ from subtitle_utils import build_cues, chunks_to_words, format_duration, to_json
|
|
| 30 |
|
| 31 |
|
| 32 |
APP_TITLE = "CaptionCraft AI"
|
| 33 |
-
MODEL_ID =
|
| 34 |
MAX_UPLOAD_MB = int(os.getenv("MAX_UPLOAD_MB", "100"))
|
| 35 |
-
MAX_AUDIO_MINUTES = int(os.getenv("MAX_AUDIO_MINUTES", "
|
| 36 |
OUTPUT_DIR = Path(tempfile.gettempdir()) / "captioncraft_downloads"
|
| 37 |
OUTPUT_DIR.mkdir(parents=True, exist_ok=True)
|
| 38 |
|
|
@@ -188,9 +188,10 @@ def _estimate_gpu_duration(audio_file: str | None = None, *args, **kwargs) -> in
|
|
| 188 |
except Exception:
|
| 189 |
# Keep the scheduler estimate conservative when metadata cannot be probed.
|
| 190 |
return 60
|
| 191 |
-
# Whisper-
|
| 192 |
-
#
|
| 193 |
-
|
|
|
|
| 194 |
|
| 195 |
|
| 196 |
@spaces.GPU(duration=_estimate_gpu_duration)
|
|
@@ -233,12 +234,12 @@ def generate_subtitles(
|
|
| 233 |
progress(0.08, desc="Preparing lightweight 16 kHz mono audio…")
|
| 234 |
_prepare_audio(source, wav_path, duration)
|
| 235 |
|
| 236 |
-
progress(0.15, desc="Whisper-
|
| 237 |
asr = _ASR_PIPELINE
|
| 238 |
|
| 239 |
generate_kwargs: dict[str, Any] = {
|
| 240 |
"task": "translate" if task_mode == "translate" else "transcribe",
|
| 241 |
-
"num_beams":
|
| 242 |
"condition_on_prev_tokens": False,
|
| 243 |
}
|
| 244 |
if language and language != "auto":
|
|
@@ -382,7 +383,7 @@ with gr.Blocks(
|
|
| 382 |
gr.HTML(
|
| 383 |
"""
|
| 384 |
<section id="cc-hero">
|
| 385 |
-
<span class="cc-chip">WHISPER-
|
| 386 |
<h1>CaptionCraft AI</h1>
|
| 387 |
<p>Turn spoken audio into clean, timed subtitle files. ZeroGPU is allocated only for transcription. Download SRT, VTT, TXT or JSON — no video rendering or burned-in captions.</p>
|
| 388 |
</section>
|
|
@@ -409,7 +410,7 @@ with gr.Blocks(
|
|
| 409 |
choices=LANGUAGES,
|
| 410 |
value="auto",
|
| 411 |
label="Spoken language",
|
| 412 |
-
info="Auto-detect is recommended. Whisper-
|
| 413 |
scale=1,
|
| 414 |
)
|
| 415 |
task_input = gr.Radio(
|
|
@@ -449,9 +450,9 @@ with gr.Blocks(
|
|
| 449 |
with gr.Accordion("How it works & limitations", open=False):
|
| 450 |
gr.Markdown(
|
| 451 |
"""
|
| 452 |
-
- **Model:** `openai/whisper-
|
| 453 |
-
- **Execution:** Hugging Face ZeroGPU. Whisper-
|
| 454 |
-
- **Resource use:** Audio is converted to mono 16 kHz WAV, then transcribed in overlapping 30-second chunks using FP16 and PyTorch inference mode.
|
| 455 |
- **Free quota:** Hugging Face documents 5 minutes of ZeroGPU time per day for free accounts. The app estimates requested GPU duration from the uploaded audio length; actual availability depends on remaining daily quota.
|
| 456 |
- **Timing:** Whisper word timestamps are used where available. If the model/runtime cannot provide them, segment timestamps are used and individual word timings are proportionally estimated.
|
| 457 |
- **No video editing:** This tool only creates caption files. It never renders a video or burns subtitles into a video.
|
|
@@ -459,7 +460,7 @@ with gr.Blocks(
|
|
| 459 |
"""
|
| 460 |
)
|
| 461 |
|
| 462 |
-
gr.HTML("<div class='cc-foot'>CaptionCraft AI · OpenAI Whisper-
|
| 463 |
|
| 464 |
generate_btn.click(
|
| 465 |
fn=generate_subtitles,
|
|
|
|
| 30 |
|
| 31 |
|
| 32 |
APP_TITLE = "CaptionCraft AI"
|
| 33 |
+
MODEL_ID = "openai/whisper-medium"
|
| 34 |
MAX_UPLOAD_MB = int(os.getenv("MAX_UPLOAD_MB", "100"))
|
| 35 |
+
MAX_AUDIO_MINUTES = int(os.getenv("MAX_AUDIO_MINUTES", "20"))
|
| 36 |
OUTPUT_DIR = Path(tempfile.gettempdir()) / "captioncraft_downloads"
|
| 37 |
OUTPUT_DIR.mkdir(parents=True, exist_ok=True)
|
| 38 |
|
|
|
|
| 188 |
except Exception:
|
| 189 |
# Keep the scheduler estimate conservative when metadata cannot be probed.
|
| 190 |
return 60
|
| 191 |
+
# Whisper-medium with beam search needs a larger reservation than small.
|
| 192 |
+
# Keep every request below the 240-second per-call cap; ZeroGPU still
|
| 193 |
+
# checks the visitor's remaining daily quota before allocating the GPU.
|
| 194 |
+
return max(30, min(240, int(math.ceil(30 + audio_seconds * 0.12))))
|
| 195 |
|
| 196 |
|
| 197 |
@spaces.GPU(duration=_estimate_gpu_duration)
|
|
|
|
| 234 |
progress(0.08, desc="Preparing lightweight 16 kHz mono audio…")
|
| 235 |
_prepare_audio(source, wav_path, duration)
|
| 236 |
|
| 237 |
+
progress(0.15, desc="Whisper-medium is ready; reserving ZeroGPU for transcription…")
|
| 238 |
asr = _ASR_PIPELINE
|
| 239 |
|
| 240 |
generate_kwargs: dict[str, Any] = {
|
| 241 |
"task": "translate" if task_mode == "translate" else "transcribe",
|
| 242 |
+
"num_beams": 5,
|
| 243 |
"condition_on_prev_tokens": False,
|
| 244 |
}
|
| 245 |
if language and language != "auto":
|
|
|
|
| 383 |
gr.HTML(
|
| 384 |
"""
|
| 385 |
<section id="cc-hero">
|
| 386 |
+
<span class="cc-chip">WHISPER-MEDIUM · ZERO GPU</span><span class="cc-chip">AUDIO → SUBTITLE FILES</span>
|
| 387 |
<h1>CaptionCraft AI</h1>
|
| 388 |
<p>Turn spoken audio into clean, timed subtitle files. ZeroGPU is allocated only for transcription. Download SRT, VTT, TXT or JSON — no video rendering or burned-in captions.</p>
|
| 389 |
</section>
|
|
|
|
| 410 |
choices=LANGUAGES,
|
| 411 |
value="auto",
|
| 412 |
label="Spoken language",
|
| 413 |
+
info="Auto-detect is recommended. Whisper-medium supports 99 languages; the dropdown lists common languages.",
|
| 414 |
scale=1,
|
| 415 |
)
|
| 416 |
task_input = gr.Radio(
|
|
|
|
| 450 |
with gr.Accordion("How it works & limitations", open=False):
|
| 451 |
gr.Markdown(
|
| 452 |
"""
|
| 453 |
+
- **Model:** `openai/whisper-medium` from Hugging Face, a multilingual 769M-parameter ASR model.
|
| 454 |
+
- **Execution:** Hugging Face ZeroGPU. Whisper-medium is loaded onto the shared CUDA device at startup and the transcription callback is wrapped with `@spaces.GPU`.
|
| 455 |
+
- **Resource use:** Audio is converted to mono 16 kHz WAV, then transcribed in overlapping 30-second chunks using FP16, 5-beam decoding and PyTorch inference mode.
|
| 456 |
- **Free quota:** Hugging Face documents 5 minutes of ZeroGPU time per day for free accounts. The app estimates requested GPU duration from the uploaded audio length; actual availability depends on remaining daily quota.
|
| 457 |
- **Timing:** Whisper word timestamps are used where available. If the model/runtime cannot provide them, segment timestamps are used and individual word timings are proportionally estimated.
|
| 458 |
- **No video editing:** This tool only creates caption files. It never renders a video or burns subtitles into a video.
|
|
|
|
| 460 |
"""
|
| 461 |
)
|
| 462 |
|
| 463 |
+
gr.HTML("<div class='cc-foot'>CaptionCraft AI · OpenAI Whisper-medium · Apache-2.0 model license</div>")
|
| 464 |
|
| 465 |
generate_btn.click(
|
| 466 |
fn=generate_subtitles,
|
tests/test_zerogpu_config.py
CHANGED
|
@@ -33,3 +33,11 @@ def test_app_uses_zero_gpu_and_does_not_mask_cuda():
|
|
| 33 |
assert "torch.float16" in app
|
| 34 |
assert "dtype=MODEL_DTYPE" in app
|
| 35 |
assert "torch_dtype=MODEL_DTYPE" not in app
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 33 |
assert "torch.float16" in app
|
| 34 |
assert "dtype=MODEL_DTYPE" in app
|
| 35 |
assert "torch_dtype=MODEL_DTYPE" not in app
|
| 36 |
+
|
| 37 |
+
|
| 38 |
+
def test_app_uses_whisper_medium_and_beam_search_for_quality():
|
| 39 |
+
app = (ROOT / "app.py").read_text(encoding="utf-8")
|
| 40 |
+
assert 'MODEL_ID = "openai/whisper-medium"' in app
|
| 41 |
+
assert '"num_beams": 5' in app
|
| 42 |
+
assert "769M-parameter" in app
|
| 43 |
+
assert "WHISPER-SMALL" not in app
|