Create app.py
Browse files
app.py
ADDED
|
@@ -0,0 +1,102 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import streamlit as st
|
| 2 |
+
import subprocess
|
| 3 |
+
import whisper
|
| 4 |
+
import json
|
| 5 |
+
import os
|
| 6 |
+
import requests
|
| 7 |
+
from datetime import timedelta
|
| 8 |
+
|
| 9 |
+
# --- Configuration ---
|
| 10 |
+
# This key should be set in your Hugging Face Space Secrets
|
| 11 |
+
API_KEY = os.environ.get("GEMINI_API_KEY", "")
|
| 12 |
+
|
| 13 |
+
def format_timecode(seconds, fps=25):
|
| 14 |
+
td = timedelta(seconds=seconds)
|
| 15 |
+
total_seconds = int(td.total_seconds())
|
| 16 |
+
hours = total_seconds // 3600
|
| 17 |
+
minutes = (total_seconds % 3600) // 60
|
| 18 |
+
secs = total_seconds % 60
|
| 19 |
+
frames = int((seconds - total_seconds) * fps)
|
| 20 |
+
return f"{hours:02}:{minutes:02}:{secs:02}:{frames:02}"
|
| 21 |
+
|
| 22 |
+
def generate_cmx_edl(edl_title, segments, source_name, fps=25):
|
| 23 |
+
edl_lines = [f"TITLE: {edl_title}", "FCM: NON-DROP FRAME\n"]
|
| 24 |
+
rec_start = 0.0
|
| 25 |
+
for i, seg in enumerate(segments, 1):
|
| 26 |
+
src_in = format_timecode(seg['src_start'], fps)
|
| 27 |
+
src_out = format_timecode(seg['src_end'], fps)
|
| 28 |
+
duration = seg['src_end'] - seg['src_start']
|
| 29 |
+
rec_in = format_timecode(rec_start, fps)
|
| 30 |
+
rec_out = format_timecode(rec_start + duration, fps)
|
| 31 |
+
edl_lines.append(f"{i:03} AX V C {src_in} {src_out} {rec_in} {rec_out}")
|
| 32 |
+
edl_lines.append(f"* FROM CLIP NAME: {source_name}")
|
| 33 |
+
edl_lines.append(f"* {seg.get('note', 'Clip')}\n")
|
| 34 |
+
rec_start += duration
|
| 35 |
+
return "\n".join(edl_lines)
|
| 36 |
+
|
| 37 |
+
def call_gemini_for_edl(transcript_data, story_prompt):
|
| 38 |
+
if not API_KEY:
|
| 39 |
+
st.error("Gemini API Key not found in Environment Variables. Please set GEMINI_API_KEY in Space Secrets.")
|
| 40 |
+
return None
|
| 41 |
+
url = f"https://generativelanguage.googleapis.com/v1beta/models/gemini-2.5-flash-preview-09-2025:generateContent?key={API_KEY}"
|
| 42 |
+
system_prompt = (
|
| 43 |
+
"You are an expert Documentary Senior Editor. Use the transcript JSON to create a condensed story. "
|
| 44 |
+
"Output ONLY a valid JSON array of segments with 'src_start', 'src_end', and 'note'. "
|
| 45 |
+
"Remove fluff, repeats, and interviewer interruptions. Focus on the most engaging content."
|
| 46 |
+
)
|
| 47 |
+
payload = {
|
| 48 |
+
"contents": [{"parts": [{"text": f"Brief: {story_prompt}\n\nTranscript:\n{json.dumps(transcript_data)}"}]}],
|
| 49 |
+
"systemInstruction": {"parts": [{"text": system_prompt}]},
|
| 50 |
+
"generationConfig": {"responseMimeType": "application/json"}
|
| 51 |
+
}
|
| 52 |
+
try:
|
| 53 |
+
res = requests.post(url, json=payload)
|
| 54 |
+
res.raise_for_status()
|
| 55 |
+
return json.loads(res.json()['candidates'][0]['content']['parts'][0]['text'])
|
| 56 |
+
except Exception as e:
|
| 57 |
+
st.error(f"Error calling Gemini: {e}")
|
| 58 |
+
return None
|
| 59 |
+
|
| 60 |
+
# --- App Interface ---
|
| 61 |
+
st.set_page_config(page_title="DocAI Editor", layout="wide")
|
| 62 |
+
st.title("Documentary AI: Senior Editor Pipeline")
|
| 63 |
+
|
| 64 |
+
uploaded_file = st.file_uploader("Upload Video/Audio", type=["mp4", "m4a", "wav", "mp3", "mov"])
|
| 65 |
+
fps = st.sidebar.number_input("Timeline FPS", value=25)
|
| 66 |
+
|
| 67 |
+
if uploaded_file:
|
| 68 |
+
if "transcript" not in st.session_state:
|
| 69 |
+
if st.button("Step 1: Transcribe"):
|
| 70 |
+
with st.spinner("Processing... Stripping audio and transcribing (this may take a few minutes):"):
|
| 71 |
+
# Save uploaded file
|
| 72 |
+
with open("temp_input", "wb") as f:
|
| 73 |
+
f.write(uploaded_file.getbuffer())
|
| 74 |
+
|
| 75 |
+
# Extract audio using FFmpeg
|
| 76 |
+
subprocess.run([
|
| 77 |
+
"ffmpeg", "-i", "temp_input", "-vn", "-acodec", "pcm_s16le",
|
| 78 |
+
"-ar", "44100", "-ac", "2", "audio.wav", "-y"
|
| 79 |
+
])
|
| 80 |
+
|
| 81 |
+
# Transcribe with Whisper
|
| 82 |
+
model = whisper.load_model("base")
|
| 83 |
+
result = model.transcribe("audio.wav", word_timestamps=True)
|
| 84 |
+
st.session_state.transcript = result['segments']
|
| 85 |
+
st.success("Transcription Complete.")
|
| 86 |
+
|
| 87 |
+
if "transcript" in st.session_state:
|
| 88 |
+
st.divider()
|
| 89 |
+
brief = st.text_area("Creative Brief", placeholder="e.g., Create a 2-minute edit focusing on the 'uncorking the brain' story and the importance of public education.")
|
| 90 |
+
if st.button("Step 2: Create EDL"):
|
| 91 |
+
with st.spinner("Senior Editor analyzing transcript..."):
|
| 92 |
+
segments = call_gemini_for_edl(st.session_state.transcript, brief)
|
| 93 |
+
if segments:
|
| 94 |
+
edl_content = generate_cmx_edl("AI_Edit", segments, uploaded_file.name, fps)
|
| 95 |
+
st.subheader("EDL Preview")
|
| 96 |
+
st.text_area("Raw CMX 3600 Output", value=edl_content, height=250)
|
| 97 |
+
st.download_button(
|
| 98 |
+
label="Download EDL for Resolve",
|
| 99 |
+
data=edl_content,
|
| 100 |
+
file_name=f"{os.path.splitext(uploaded_file.name)[0]}_AI_Edit.edl",
|
| 101 |
+
mime="text/plain"
|
| 102 |
+
)
|