speechanalytics / app.py
Waqas167's picture
Update app.py
5b28093 verified
raw
history blame
10.6 kB
# ---- BEGIN: HF Spaces permission fix (must be first!) ----
import os, pathlib
HOME_DIR = "/home/user" # writable on Hugging Face Spaces
os.environ["HOME"] = HOME_DIR
os.environ["XDG_CONFIG_HOME"] = HOME_DIR
CONFIG_DIR = os.path.join(HOME_DIR, ".streamlit")
os.makedirs(CONFIG_DIR, exist_ok=True)
# Tell Streamlit where to read/write config & metrics
os.environ["STREAMLIT_CONFIG_DIR"] = CONFIG_DIR
os.environ["STREAMLIT_BROWSER_GATHERUSAGESTATS"] = "false"
# ---- END: HF Spaces permission fix ----
import io, numpy as np, librosa, torch, soundfile as sf
from transformers import AutoProcessor, Wav2Vec2ForCTC
from pydub import AudioSegment
from moviepy.editor import VideoFileClip
from google import genai
from google.genai import types
from streamlit_mic_recorder import mic_recorder
import streamlit as st # <-- import AFTER the env vars above
# ---------------- Config ----------------
st.set_page_config(page_title="Urdu Speech Analyzer", page_icon="πŸŽ™οΈ", layout="wide")
PAGE_TITLE = "πŸŽ™οΈ Urdu Audio & Video Speech Analyzer"
model_id = "facebook/mms-1b-l1107"
lang_code = "urd-script_arabic"
api_key = "AIzaSyBEWWn32PxVEaUsoe67GJOEpF4FQT87Kxo" # hard-coded as requested
# ---------------- Model ----------------
@st.cache_resource
def load_model_and_processor():
processor = AutoProcessor.from_pretrained(model_id, target_lang=lang_code)
model = Wav2Vec2ForCTC.from_pretrained(
model_id, target_lang=lang_code, ignore_mismatched_sizes=True
)
model.load_adapter(lang_code)
return processor, model
processor, model = load_model_and_processor()
# ---------------- Helpers ----------------
def get_wav_from_input(file_path, output_path="converted.wav"):
ext = os.path.splitext(file_path)[-1].lower()
if ext in [".mp4", ".mkv", ".avi", ".mov"]:
video = VideoFileClip(file_path)
video.audio.write_audiofile(output_path, fps=16000)
elif ext in [".mp3", ".aac", ".flac", ".ogg", ".m4a"]:
audio = AudioSegment.from_file(file_path)
audio = audio.set_frame_rate(16000).set_channels(1)
audio.export(output_path, format="wav")
elif ext == ".wav":
audio = AudioSegment.from_wav(file_path)
audio = audio.set_frame_rate(16000).set_channels(1)
audio.export(output_path, format="wav")
else:
raise ValueError("Unsupported file format.")
return output_path
def save_wav_resampled(audio_f32: np.ndarray, sr_in: int, path: str):
if sr_in != 16000:
audio_f32 = librosa.resample(audio_f32, orig_sr=sr_in, target_sr=16000)
audio_f32 = librosa.util.normalize(audio_f32)
sf.write(path, audio_f32.astype(np.float32), 16000)
def transcribe(wav_path) -> str:
audio, sr = librosa.load(wav_path, sr=16000, mono=True)
inputs = processor(audio, sampling_rate=sr, return_tensors="pt", padding=True)
with torch.no_grad():
logits = model(**inputs).logits
pred_ids = torch.argmax(logits, dim=-1)
return processor.batch_decode(pred_ids)[0]
def analyze_transcript(transcript: str) -> str:
client = genai.Client(api_key=api_key)
system_instr = """
You are a speech analyst. The following transcription is in Urdu and contains no punctuation β€” your first task is to correct the transcript by segmenting it into grammatically correct sentences.
Then:
1. Translate the corrected Urdu transcript into English.
2. Determine whether the transcript involves a single speaker or multiple speakers.
3. If multiple speakers are detected, perform diarization by segmenting the transcript with clear speaker labels.
⚠️ Format the segmented transcript *exactly* like this:
**Segmented Transcript**
**Urdu:**
Person 01:
[Urdu line here]
Person 02:
[Urdu line here]
...
**English:**
Person 01:
[English line here]
Person 02:
[English line here]
...
After that, provide your analysis in the following format:
**Speaker-wise Analysis**
[One or two sentences per speaker about tone, emotion, behavior]
**Sentiment and Communication Style**
[Concise overall tone: e.g., friendly, formal, tense, etc.]
**Summary of Discussion**
[A 2–3 line summary of what the speakers talked about, in English]
"""
resp = client.models.generate_content(
model="gemini-2.5-flash",
contents=[transcript],
config=types.GenerateContentConfig(system_instruction=system_instr, temperature=0.0)
)
return resp.text
def format_transcript_block(text: str) -> str:
lines = text.split("Person ")
out = ""
for line in lines:
line = line.strip()
if not line:
continue
if line.startswith("01:") or line.startswith("02:"):
out += f"\n**Person {line[:2]}**:\n{line[3:].strip()}\n\n"
else:
out += f"{line}\n\n"
return out
# ---------------- Header ----------------
st.markdown(f"""
<div style="text-align: left; padding-bottom: 1rem;">
<h1 style='color:#1f77b4; font-size: 2.5em; font-weight: 800; margin-bottom: 0.2em;'>
{PAGE_TITLE}
</h1>
<p style='color: #7c8a98; font-size: 1.05em; margin-top: 0;'>
Record or upload Urdu speech for structured transcription, diarization, and smart AI analysis.
</p>
</div>
""", unsafe_allow_html=True)
# ================= Mic: true Start/Stop + narrow Analyze =================
st.markdown("### 🎀 Live recording")
# The component renders **Start** and **Stop** buttons and keeps recording until you press Stop.
rec = mic_recorder(
start_prompt="▢️ Start",
stop_prompt="⏹️ Stop",
just_once=False, # allow multiple recordings in a session
key="recorder",
format="wav" # returns WAV bytes
)
# `rec` returns after Stop. Different versions return bytes or a dict β€” handle both.
audio_bytes, sr_in = None, 44100
if rec is not None:
if isinstance(rec, dict) and "bytes" in rec:
audio_bytes = rec["bytes"]
sr_in = int(rec.get("sample_rate", 44100))
elif isinstance(rec, (bytes, bytearray)):
audio_bytes = rec
sr_in = 44100 # component default
else:
# fallback: try to extract .get("audio") etc if lib changes
audio_bytes = rec.get("audio") if isinstance(rec, dict) else None
if audio_bytes:
st.success("Audio captured.")
# Convert to mono float32
data, sr_read = sf.read(io.BytesIO(audio_bytes), dtype="float32", always_2d=False)
if data.ndim > 1:
data = data.mean(axis=1)
if sr_read: # prefer the rate embedded in the WAV
sr_in = sr_read
# Save as 16 kHz mono for the model
tmp_wav = "mic_recording.wav"
save_wav_resampled(data, sr_in, tmp_wav)
# Minimal playback (no waveform)
st.audio(audio_bytes, format="audio/wav")
st.caption(f"Duration: {data.size / sr_in:.2f} s")
# Slim Analyze button (not full width)
if st.button("πŸ” Analyze", type="primary"):
with st.spinner("⏳ Transcribing & analyzing..."):
transcript = transcribe(tmp_wav) # raw not displayed
report = analyze_transcript(transcript)
segmented_urdu = segmented_english = analysis_only = ""
if "Urdu:" in report and "English:" in report:
u0 = report.find("Urdu:")
e0 = report.find("English:")
segmented_urdu = report[u0 + len("Urdu:"):e0].strip()
english_section = report[e0 + len("English:"):].strip()
if "**Speaker-wise Analysis**" in english_section:
parts = english_section.split("**Speaker-wise Analysis**")
segmented_english = parts[0].strip()
analysis_only = "**Speaker-wise Analysis**" + parts[1].strip()
else:
segmented_english = english_section.strip()
analysis_only = "⚠️ Could not extract structured analysis."
if segmented_urdu or segmented_english:
st.markdown("### πŸ—£οΈ Segmented Transcript")
c1, c2 = st.columns(2)
with c1:
st.markdown("#### Urdu")
st.markdown(format_transcript_block(segmented_urdu) if segmented_urdu else "_(none)_")
with c2:
st.markdown("#### English")
st.markdown(format_transcript_block(segmented_english) if segmented_english else "_(none)_")
if analysis_only:
st.markdown("### 🧠 Gemini Analysis Summary")
st.markdown(analysis_only)
st.markdown("---")
# ================= Upload (unchanged) =================
st.markdown("### πŸ“‚ Or upload an audio/video file")
uploaded_file = st.file_uploader(
label="",
type=["mp3", "mp4", "wav", "mkv", "aac", "ogg", "m4a", "flac"],
label_visibility="collapsed"
)
if uploaded_file is not None:
with st.spinner("⏳ Transcribing..."):
file_name = uploaded_file.name
temp_path = f"temp_input{os.path.splitext(file_name)[-1]}"
with open(temp_path, "wb") as f:
f.write(uploaded_file.read())
wav_path = get_wav_from_input(temp_path)
transcript = transcribe(wav_path)
with st.spinner("πŸ” Analyzing with Gemini..."):
report = analyze_transcript(transcript)
segmented_urdu = segmented_english = analysis_only = ""
if "Urdu:" in report and "English:" in report:
u0 = report.find("Urdu:")
e0 = report.find("English:")
segmented_urdu = report[u0 + len("Urdu:"):e0].strip()
english_section = report[e0 + len("English:"):].strip()
if "**Speaker-wise Analysis**" in english_section:
parts = english_section.split("**Speaker-wise Analysis**")
segmented_english = parts[0].strip()
analysis_only = "**Speaker-wise Analysis**" + parts[1].strip()
else:
segmented_english = english_section.strip()
analysis_only = "⚠️ Could not extract structured analysis."
if segmented_urdu or segmented_english:
st.markdown("### πŸ—£οΈ Segmented Transcript")
c1, c2 = st.columns(2)
with c1:
st.markdown("#### Urdu")
st.markdown(format_transcript_block(segmented_urdu) if segmented_urdu else "_(none)_")
with c2:
st.markdown("#### English")
st.markdown(format_transcript_block(segmented_english) if segmented_english else "_(none)_")
if analysis_only:
st.markdown("### 🧠 Gemini Analysis Summary")
st.markdown(analysis_only)