ALLARD Marc-Antoine
d
316846a
raw
history blame
25.7 kB
import streamlit as st
import os
import json
import wave
import numpy as np
from datetime import timedelta
import base64
from io import BytesIO, StringIO
import tempfile
# Page configuration
st.set_page_config(
page_title="ASR Annotation Tool",
page_icon="🎀",
layout="wide",
initial_sidebar_state="expanded"
)
# Initialize session state
if 'annotation_type' not in st.session_state:
st.session_state.annotation_type = None
if 'audio_file' not in st.session_state:
st.session_state.audio_file = None
if 'transcript' not in st.session_state:
st.session_state.transcript = ""
if 'segments' not in st.session_state:
st.session_state.segments = []
if 'current_page' not in st.session_state:
st.session_state.current_page = "home"
if 'audio_duration' not in st.session_state:
st.session_state.audio_duration = 0
def get_audio_duration(audio_file):
"""Get audio duration in seconds"""
try:
with wave.open(audio_file, 'rb') as wav_file:
frames = wav_file.getnframes()
sample_rate = wav_file.getframerate()
duration = frames / float(sample_rate)
return duration
except:
return 0
def format_time(seconds):
"""Format seconds to HH:MM:SS.mmm"""
td = timedelta(seconds=seconds)
total_seconds = int(td.total_seconds())
hours, remainder = divmod(total_seconds, 3600)
minutes, seconds = divmod(remainder, 60)
milliseconds = int((td.total_seconds() - total_seconds) * 1000)
return f"{hours:02d}:{minutes:02d}:{seconds:02d}.{milliseconds:03d}"
def create_audio_player_html(audio_data, audio_id="audio_player"):
"""Create HTML audio player with controls"""
audio_base64 = base64.b64encode(audio_data).decode()
html = f"""
<div style="margin: 20px 0;">
<audio id="{audio_id}" controls style="width: 100%; height: 40px;">
<source src="data:audio/wav;base64,{audio_base64}" type="audio/wav">
Your browser does not support the audio element.
</audio>
<div style="margin-top: 10px;">
<button onclick="document.getElementById('{audio_id}').currentTime -= 5"
style="margin-right: 5px; padding: 5px 10px; background: #ff4b4b; color: white; border: none; border-radius: 3px; cursor: pointer;">
βͺ -5s
</button>
<button onclick="document.getElementById('{audio_id}').currentTime -= 1"
style="margin-right: 5px; padding: 5px 10px; background: #ff4b4b; color: white; border: none; border-radius: 3px; cursor: pointer;">
βͺ -1s
</button>
<button onclick="var audio = document.getElementById('{audio_id}'); audio.paused ? audio.play() : audio.pause()"
style="margin-right: 5px; padding: 5px 15px; background: #00cc44; color: white; border: none; border-radius: 3px; cursor: pointer;">
⏯️ Play/Pause
</button>
<button onclick="document.getElementById('{audio_id}').currentTime += 1"
style="margin-right: 5px; padding: 5px 10px; background: #ff4b4b; color: white; border: none; border-radius: 3px; cursor: pointer;">
+1s ⏩
</button>
<button onclick="document.getElementById('{audio_id}').currentTime += 5"
style="padding: 5px 10px; background: #ff4b4b; color: white; border: none; border-radius: 3px; cursor: pointer;">
+5s ⏩
</button>
</div>
</div>
"""
return html
def create_waveform_html(audio_data, segments=None):
"""Create interactive waveform with region selection"""
audio_base64 = base64.b64encode(audio_data).decode()
segments_json = json.dumps(segments or [])
html = f"""
<div id="waveform-container" style="margin: 20px 0;">
<div id="waveform" style="height: 200px; border: 1px solid #ddd;"></div>
<div style="margin-top: 10px;">
<button id="play-pause" style="margin-right: 5px; padding: 8px 15px; background: #00cc44; color: white; border: none; border-radius: 3px; cursor: pointer;">
⏯️ Play/Pause
</button>
<button id="add-region" style="margin-right: 5px; padding: 8px 15px; background: #0066cc; color: white; border: none; border-radius: 3px; cursor: pointer;">
βž• Add Region
</button>
<button id="clear-regions" style="padding: 8px 15px; background: #cc0000; color: white; border: none; border-radius: 3px; cursor: pointer;">
πŸ—‘οΈ Clear All
</button>
</div>
<div id="regions-list" style="margin-top: 15px; max-height: 200px; overflow-y: auto; color: white;">
<h4>Segments:</h4>
<div id="segments-container"></div>
</div>
</div>
<script src="https://cdnjs.cloudflare.com/ajax/libs/wavesurfer.js/6.6.4/wavesurfer.min.js"></script>
<script src="https://cdnjs.cloudflare.com/ajax/libs/wavesurfer.js/6.6.4/plugin/wavesurfer.regions.min.js"></script>
<script>
let wavesurfer;
let regions = {segments_json};
let speakerColors = {{}};
// Initialize WaveSurfer
wavesurfer = WaveSurfer.create({{
container: '#waveform',
waveColor: '#4FC3F7',
progressColor: '#1976D2',
height: 200,
responsive: true,
plugins: [
WaveSurfer.regions.create({{
dragSelection: true,
color: 'rgba(255, 75, 75, 0.3)'
}})
]
}});
// Load audio
wavesurfer.load('data:audio/wav;base64,{audio_base64}');
// Play/Pause button
document.getElementById('play-pause').addEventListener('click', function() {{
wavesurfer.playPause();
}});
// Add region button
document.getElementById('add-region').addEventListener('click', function() {{
const start = wavesurfer.getCurrentTime();
const end = Math.min(start + 2, wavesurfer.getDuration());
// Ask for speaker ID first
const speakerId = prompt("Enter speaker ID (e.g., SPK001):", "SPK" + (Object.keys(speakerColors).length + 1).toString().padStart(3, '0'));
if (speakerId) {{
addRegion(start, end, speakerId);
}}
}});
// Clear regions button
document.getElementById('clear-regions').addEventListener('click', function() {{
wavesurfer.clearRegions();
regions = [];
updateRegionsList();
}});
// Add region function
function addRegion(start, end, speaker_id) {{
// Get or assign color for this speaker
if (!speakerColors[speaker_id]) {{
speakerColors[speaker_id] = getColorForSpeaker(speaker_id);
}}
const region = wavesurfer.addRegion({{
start: start,
end: end,
color: speakerColors[speaker_id],
drag: true,
resize: true
}});
regions.push({{
id: region.id,
start: start,
end: end,
speaker_id: speaker_id
}});
updateRegionsList();
}}
// Update regions list
function updateRegionsList() {{
const container = document.getElementById('segments-container');
container.innerHTML = '';
regions.forEach((region, index) => {{
const div = document.createElement('div');
div.style.cssText = 'border: 1px solid #ddd; padding: 10px; margin: 5px 0; border-radius: 5px; color: white;';
div.innerHTML = `
<div style="display: flex; justify-content: space-between; align-items: center;">
<div>
<strong>Segment ${{index + 1}}</strong><br>
Start: ${{region.start.toFixed(2)}}s | End: ${{region.end.toFixed(2)}}s<br>
<input type="text" value="${{region.speaker_id}}"
onchange="updateSpeakerId('${{region.id}}', this.value, '${{region.speaker_id}}')"
style="margin-top: 5px; padding: 3px; border: 1px solid #ccc; border-radius: 3px; color: black;">
</div>
<button onclick="removeRegion('${{region.id}}')"
style="background: #cc0000; color: white; border: none; border-radius: 3px; padding: 5px 8px; cursor: pointer;">
βœ•
</button>
</div>
`;
container.appendChild(div);
}});
}}
// Remove region
function removeRegion(regionId) {{
wavesurfer.regions.list[regionId].remove();
regions = regions.filter(r => r.id !== regionId);
updateRegionsList();
}}
// Update speaker ID
function updateSpeakerId(regionId, newId, oldId) {{
const region = regions.find(r => r.id === regionId);
if (region) {{
region.speaker_id = newId;
// Update color if speaker ID changed
if (newId !== oldId) {{
if (!speakerColors[newId]) {{
speakerColors[newId] = getColorForSpeaker(newId);
}}
wavesurfer.regions.list[regionId].color = speakerColors[newId];
wavesurfer.regions.list[regionId].updateRender();
}}
}}
}}
// Get consistent color for a specific speaker
function getColorForSpeaker(speakerId) {{
const colors = [
'rgba(255, 75, 75, 0.3)', // Red
'rgba(75, 192, 75, 0.3)', // Green
'rgba(75, 75, 255, 0.3)', // Blue
'rgba(255, 192, 75, 0.3)', // Yellow
'rgba(255, 75, 255, 0.3)', // Magenta
'rgba(75, 192, 192, 0.3)', // Cyan
'rgba(192, 75, 192, 0.3)', // Purple
'rgba(192, 192, 75, 0.3)' // Olive
];
// Generate a deterministic index based on the speaker ID string
let hash = 0;
for (let i = 0; i < speakerId.length; i++) {{
hash = ((hash << 5) - hash) + speakerId.charCodeAt(i);
hash |= 0; // Convert to 32bit integer
}}
// Use the absolute value of hash to select a color
const index = Math.abs(hash) % colors.length;
return colors[index];
}}
// Update region on change
wavesurfer.on('region-update-end', function(region) {{
const regionData = regions.find(r => r.id === region.id);
if (regionData) {{
regionData.start = region.start;
regionData.end = region.end;
updateRegionsList();
}}
}});
// Load existing regions
wavesurfer.on('ready', function() {{
// First, create color mappings for existing speakers
regions.forEach(regionData => {{
if (!speakerColors[regionData.speaker_id]) {{
speakerColors[regionData.speaker_id] = getColorForSpeaker(regionData.speaker_id);
}}
}});
// Then create the regions with their colors
regions.forEach(regionData => {{
const region = wavesurfer.addRegion({{
start: regionData.start,
end: regionData.end,
color: speakerColors[regionData.speaker_id],
drag: true,
resize: true
}});
regionData.id = region.id;
}});
updateRegionsList();
}});
// Export regions function for Streamlit
window.getRegions = function() {{
return regions.map(r => ({{
start: r.start,
end: r.end,
speaker_id: r.speaker_id
}}));
}}
</script>
"""
return html
def generate_srt(segments, transcript):
"""Generate SRT format from segments and transcript"""
srt_content = ""
for i, segment in enumerate(segments):
start_time = format_srt_time(segment['start'])
end_time = format_srt_time(segment['end'])
# Extract corresponding text (simplified - in real app you'd need better text matching)
text = f"{segment['speaker_id']}: [Segment {i+1} text]"
srt_content += f"{i+1}\n"
srt_content += f"{start_time} --> {end_time}\n"
srt_content += f"{text}\n\n"
return srt_content
def format_srt_time(seconds):
"""Format time for SRT format (HH:MM:SS,mmm)"""
hours = int(seconds // 3600)
minutes = int((seconds % 3600) // 60)
secs = int(seconds % 60)
millisecs = int((seconds % 1) * 1000)
return f"{hours:02d}:{minutes:02d}:{secs:02d},{millisecs:03d}"
def get_download_link(content, filename, label="Download file"):
"""Generate download link for text content"""
b64 = base64.b64encode(content.encode()).decode()
href = f'<a href="data:file/txt;base64,{b64}" download="{filename}">{label}</a>'
return href
# Main App Layout
def main():
st.title("🎀 ASR Annotation Tool")
st.markdown("Simple tool for transcribing, segmenting, and annotating audio for ASR dataset creation.")
# Sidebar for navigation and settings
with st.sidebar:
st.header("Navigation")
if st.button("🏠 Home", use_container_width=True):
st.session_state.current_page = "home"
if st.session_state.audio_file and st.session_state.annotation_type:
if st.button("πŸ“ Transcription", use_container_width=True):
st.session_state.current_page = "transcription"
if st.session_state.annotation_type == "multi_speaker" and st.session_state.transcript:
if st.button("🎯 Segmentation", use_container_width=True):
st.session_state.current_page = "segmentation"
if st.session_state.segments:
if st.button("πŸ“Š Assignment", use_container_width=True):
st.session_state.current_page = "assignment"
if st.session_state.current_page == "home":
show_home_page()
elif st.session_state.current_page == "transcription":
show_transcription_page()
elif st.session_state.current_page == "segmentation":
show_segmentation_page()
elif st.session_state.current_page == "assignment":
show_assignment_page()
def show_home_page():
"""Home page - annotation type selection and file upload"""
# Annotation type selection
st.subheader("1. Select Annotation Type")
annotation_type = st.radio(
"How many speakers are in your audio?",
["single_speaker", "multi_speaker"],
format_func=lambda x: "Single Speaker (Simple ASR)" if x == "single_speaker" else "Multi Speaker (Diarization)",
key="annotation_type_radio"
)
st.session_state.annotation_type = annotation_type
# File upload with better error handling
st.subheader("2. Upload Audio File")
# Add file size warning for Hugging Face Spaces
st.info("πŸ’‘ **Tip for Hugging Face Spaces:** Large files (>10MB) may fail to upload. Try smaller audio files or compress your audio if you encounter issues.")
uploaded_file = st.file_uploader(
"Choose an audio file",
type=['wav', 'mp3', 'flac', 'm4a'],
help="Supported formats: WAV, MP3, FLAC, M4A"
)
if uploaded_file is not None:
st.session_state.audio_file = uploaded_file.read()
# Save temporary file to get duration
with tempfile.NamedTemporaryFile(delete=False, suffix='.wav') as tmp_file:
tmp_file.write(st.session_state.audio_file)
st.session_state.audio_duration = get_audio_duration(tmp_file.name)
os.unlink(tmp_file.name)
st.success(f"βœ… Audio file uploaded successfully!")
st.info(f"Duration: {format_time(st.session_state.audio_duration)}")
# Show audio player
st.subheader("Audio Preview")
audio_html = create_audio_player_html(st.session_state.audio_file)
st.components.v1.html(audio_html, height=120)
# Continue button
if st.button("Continue to Transcription β†’", type="primary"):
st.session_state.current_page = "transcription"
st.rerun()
def show_transcription_page():
"""Transcription page - text annotation"""
st.header("πŸ“ Text Transcription")
if not st.session_state.audio_file:
st.error("Please upload an audio file first!")
return
# Audio player
st.subheader("Audio Player")
audio_html = create_audio_player_html(st.session_state.audio_file)
st.components.v1.html(audio_html, height=120)
# Transcription area
st.subheader("Transcript")
transcript = st.text_area(
"Write your transcription here:",
value=st.session_state.transcript,
height=300,
help="Check the guidelines below to help you transcribe accurately."
)
st.session_state.transcript = transcript
# Guidelines reminder
with st.expander("πŸ“‹ Transcription Guidelines"):
st.markdown("""
**Key Guidelines:**
- Transcribe exactly what is said
- Use standard punctuation and capitalization (tip: Get punctuation from natural pauses in dialogue)
- Write numbers 1-10 as words, 11+ as digits
- Ignore unclear speech or marked as [unclear] or [inaudible]
- For multi-speaker: transcribe all audible speech without identifying speakers
""")
# Action buttons
col1, col2, col3 = st.columns(3)
with col1:
if transcript.strip():
download_link = get_download_link(transcript, "transcript.txt", "πŸ’Ύ Download Transcript")
st.markdown(download_link, unsafe_allow_html=True)
else:
st.button("πŸ’Ύ Download Transcript", disabled=True)
with col2:
if st.session_state.annotation_type == "multi_speaker" and transcript.strip():
if st.button("🎯 Continue to Segmentation β†’"):
st.session_state.current_page = "segmentation"
st.rerun()
with col3:
if st.session_state.annotation_type == "single_speaker" and transcript.strip():
if st.button("βœ… Finish Annotation"):
st.balloons()
st.success("πŸŽ‰ Single speaker annotation completed!")
download_link = get_download_link(transcript, "transcript.txt", "πŸ“₯ Download Final Transcript")
st.markdown(download_link, unsafe_allow_html=True)
def show_segmentation_page():
"""Segmentation page - audio region selection"""
st.header("🎯 Audio Segmentation")
if not st.session_state.audio_file:
st.error("Please upload an audio file first!")
return
st.info("Click and drag on the waveform to create segments. Resize by dragging edges, remove with βœ• button.")
# Interactive waveform
waveform_html = create_waveform_html(st.session_state.audio_file, st.session_state.segments)
st.components.v1.html(waveform_html, height=500)
# Manual segment addition
st.subheader("Manual Segment Addition")
st.info("After having segmented the wav using our wav surfer, you can manually add segments here. Don't hesitate to replay and pause for the best results.")
col1, col2, col3, col4 = st.columns(4)
with col1:
start_time = st.number_input("Start (seconds)", min_value=0.0, max_value=st.session_state.audio_duration, step=0.1)
with col2:
end_time = st.number_input("End (seconds)", min_value=0.0, max_value=st.session_state.audio_duration, step=0.1)
with col3:
speaker_id = st.text_input("Speaker ID", value="SPK001")
with col4:
if st.button("βž• Add Segment"):
if start_time < end_time:
new_segment = {
"start": start_time,
"end": end_time,
"speaker_id": speaker_id
}
st.session_state.segments.append(new_segment)
st.success("Segment added!")
st.rerun()
else:
st.error("End time must be greater than start time!")
# Current segments display
if st.session_state.segments:
st.subheader("Current Segments")
for i, segment in enumerate(st.session_state.segments):
col1, col2 = st.columns([4, 1])
with col1:
st.write(f"**Segment {i+1}:** {segment['speaker_id']} | {segment['start']:.2f}s - {segment['end']:.2f}s")
with col2:
if st.button("πŸ—‘οΈ", key=f"remove_{i}"):
st.session_state.segments.pop(i)
st.rerun()
# Continue button
if st.session_state.segments:
if st.button("πŸ“Š Continue to Assignment β†’", type="primary"):
st.session_state.current_page = "assignment"
st.rerun()
def show_assignment_page():
"""Assignment page - text-to-segment mapping and final export"""
st.header("πŸ“Š Text-Segment Assignment")
if not st.session_state.segments:
st.error("Please create segments first!")
return
st.info("Assign portions of your text transcript to each audio segment to create the final annotation.")
# Display transcript
st.subheader("Original Transcript")
st.text_area("Reference transcript:", value=st.session_state.transcript, height=150, disabled=True)
# Segment assignment
st.subheader("Segment Text Assignment")
assigned_segments = []
for i, segment in enumerate(st.session_state.segments):
st.write(f"**Segment {i+1}:** {segment['speaker_id']} ({segment['start']:.2f}s - {segment['end']:.2f}s)")
segment_text = st.text_area(
f"Text for segment {i+1}:",
key=f"segment_text_{i}",
height=100,
help="Copy and paste the relevant portion of the text transcript for this segment"
)
assigned_segments.append({
**segment,
"text": segment_text
})
st.divider()
# Preview SRT
if st.button("πŸ” Preview SRT"):
srt_preview = generate_srt_with_text(assigned_segments)
st.subheader("SRT Preview")
st.code(srt_preview, language="text")
# Final save
st.subheader("Download Final Annotation")
col1, col2 = st.columns(2)
with col1:
# Create enhanced transcript with speaker labels
enhanced_transcript = create_speaker_transcript(assigned_segments)
download_transcript = get_download_link(enhanced_transcript, "final_transcript.txt", "πŸ’Ύ Download Transcript")
st.markdown(download_transcript, unsafe_allow_html=True)
with col2:
srt_content = generate_srt_with_text(assigned_segments)
download_srt = get_download_link(srt_content, "final_transcript.srt", "πŸ’Ύ Download SRT")
st.markdown(download_srt, unsafe_allow_html=True)
if st.button("πŸŽ‰ Finish Annotation", type="primary"):
st.balloons()
st.success("πŸŽ‰ Yihawww or Youhouuuu Multi-speaker annotation completed!")
# Final downloads
st.subheader("Download your files:")
download_transcript = get_download_link(enhanced_transcript, "final_transcript.txt", "πŸ“₯ Download Transcript")
download_srt = get_download_link(srt_content, "final_transcript.srt", "πŸ“₯ Download SRT")
st.markdown(download_transcript, unsafe_allow_html=True)
st.markdown(download_srt, unsafe_allow_html=True)
if st.button("πŸ”„ Back to Segmentation"):
st.session_state.current_page = "segmentation"
st.rerun()
def generate_srt_with_text(segments):
"""Generate SRT with actual text content"""
srt_content = ""
for i, segment in enumerate(segments):
start_time = format_srt_time(segment['start'])
end_time = format_srt_time(segment['end'])
text = segment.get('text', '').strip() or f"[Segment {i+1} - No text assigned]"
srt_content += f"{i+1}\n"
srt_content += f"{start_time} --> {end_time}\n"
srt_content += f"{segment['speaker_id']}: {text}\n\n"
return srt_content
def create_speaker_transcript(segments):
"""Create speaker-labeled transcript"""
transcript_lines = []
for segment in sorted(segments, key=lambda x: x['start']):
text = segment.get('text', '').strip()
if text:
transcript_lines.append(f"{segment['speaker_id']}: {text}")
return "\n\n".join(transcript_lines)
if __name__ == "__main__":
main()