Commit ·
256ade0
1
Parent(s): ad594d3
Added app code
Browse files- app.py +244 -0
- requirements.txt +84 -0
app.py
ADDED
|
@@ -0,0 +1,244 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import os
|
| 2 |
+
import time
|
| 3 |
+
from typing import Optional, Tuple
|
| 4 |
+
|
| 5 |
+
import gradio as gr
|
| 6 |
+
import librosa
|
| 7 |
+
import numpy as np
|
| 8 |
+
import torch
|
| 9 |
+
from transformers import WhisperForConditionalGeneration, WhisperProcessor
|
| 10 |
+
from openai import OpenAI
|
| 11 |
+
|
| 12 |
+
# -----------------------------
|
| 13 |
+
# Models / Config
|
| 14 |
+
# -----------------------------
|
| 15 |
+
BASE_REPO = "openai/whisper-large-v3"
|
| 16 |
+
|
| 17 |
+
FINETUNED_REPO = "dmatekenya/whisper-large-v3-chichewa"
|
| 18 |
+
FINETUNED_REVISION = "bff60fb08ba9f294e05bfcab4306f30b6a0cfc0a" # pinned commit hash
|
| 19 |
+
|
| 20 |
+
# Local WhisperProcessor language hint (keep consistent with how you evaluated)
|
| 21 |
+
LOCAL_LANGUAGE = "shona"
|
| 22 |
+
|
| 23 |
+
# OpenAI language hint (ISO-639-1). Chichewa/Nyanja often "ny". Set to None to auto-detect.
|
| 24 |
+
OPENAI_LANGUAGE = "ny"
|
| 25 |
+
|
| 26 |
+
# Audio constraints
|
| 27 |
+
TARGET_SR = 16000
|
| 28 |
+
MAX_SECONDS = 30.0
|
| 29 |
+
|
| 30 |
+
# OpenAI transcription model (commercial)
|
| 31 |
+
OPENAI_MODEL = "whisper-1" # simple + stable
|
| 32 |
+
|
| 33 |
+
# -----------------------------
|
| 34 |
+
# UI Text / Styling
|
| 35 |
+
# -----------------------------
|
| 36 |
+
LOGO_HTML = """
|
| 37 |
+
<div style="text-align:center; margin-bottom: 25px;">
|
| 38 |
+
<img src="https://i.ibb.co/5nQdGSs/logo.png"
|
| 39 |
+
style="max-width: 100%; height: auto; border-radius: 12px;">
|
| 40 |
+
</div>
|
| 41 |
+
"""
|
| 42 |
+
|
| 43 |
+
TITLE_HTML = """
|
| 44 |
+
<h1 style="text-align:center; font-size:34px; margin-bottom:10px;">
|
| 45 |
+
Chichewa Speech2Text: How Custom Data Improves Performance
|
| 46 |
+
</h1>
|
| 47 |
+
"""
|
| 48 |
+
|
| 49 |
+
HIGHLIGHT_TEXT = """
|
| 50 |
+
<p style="text-align:center; font-size:20px; font-weight:600; color:#1F3A5F; margin-bottom:20px;">
|
| 51 |
+
Observe how the fine-tuned model provides better transcription quality.
|
| 52 |
+
</p>
|
| 53 |
+
"""
|
| 54 |
+
|
| 55 |
+
DESCRIPTION_HTML = """
|
| 56 |
+
<p style="text-align:center; font-size:18px; margin-bottom: 18px;">
|
| 57 |
+
Upload or record a short Chichewa voice note (≤30 seconds). The same audio will be transcribed by three systems.
|
| 58 |
+
</p>
|
| 59 |
+
"""
|
| 60 |
+
|
| 61 |
+
ARTICLE_HTML = """
|
| 62 |
+
<p style="text-align:center; margin-top: 10px;">
|
| 63 |
+
Read more about the <a href="https://dmatekenya.github.io/Chichewa-Speech2Text/README.html" target="_blank">ChichewaSpeech2Text</a> project
|
| 64 |
+
and sign up for our voice note donation event:
|
| 65 |
+
<a href="https://forms.gle/fHLESutofVvb2YFM9" target="_blank">Google Form</a>.
|
| 66 |
+
</p>
|
| 67 |
+
"""
|
| 68 |
+
|
| 69 |
+
# -----------------------------
|
| 70 |
+
# Load local models once
|
| 71 |
+
# -----------------------------
|
| 72 |
+
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
|
| 73 |
+
DTYPE = torch.float16 if DEVICE == "cuda" else torch.float32
|
| 74 |
+
|
| 75 |
+
print(f"Using device: {DEVICE}", flush=True)
|
| 76 |
+
|
| 77 |
+
PROCESSOR = WhisperProcessor.from_pretrained(
|
| 78 |
+
BASE_REPO,
|
| 79 |
+
language=LOCAL_LANGUAGE,
|
| 80 |
+
task="transcribe",
|
| 81 |
+
)
|
| 82 |
+
|
| 83 |
+
MODEL_BASE = WhisperForConditionalGeneration.from_pretrained(BASE_REPO).to(DEVICE).eval()
|
| 84 |
+
MODEL_FT = WhisperForConditionalGeneration.from_pretrained(
|
| 85 |
+
FINETUNED_REPO,
|
| 86 |
+
revision=FINETUNED_REVISION,
|
| 87 |
+
).to(DEVICE).eval()
|
| 88 |
+
|
| 89 |
+
if DEVICE == "cuda":
|
| 90 |
+
MODEL_BASE = MODEL_BASE.to(dtype=DTYPE)
|
| 91 |
+
MODEL_FT = MODEL_FT.to(dtype=DTYPE)
|
| 92 |
+
|
| 93 |
+
OPENAI_CLIENT = OpenAI()
|
| 94 |
+
|
| 95 |
+
|
| 96 |
+
# -----------------------------
|
| 97 |
+
# Helpers
|
| 98 |
+
# -----------------------------
|
| 99 |
+
def load_audio(audio_path: str) -> Tuple[np.ndarray, int, float]:
|
| 100 |
+
y, sr = librosa.load(audio_path, sr=TARGET_SR, mono=True)
|
| 101 |
+
dur = float(len(y) / sr) if sr else 0.0
|
| 102 |
+
return y, sr, dur
|
| 103 |
+
|
| 104 |
+
|
| 105 |
+
@torch.inference_mode()
|
| 106 |
+
def transcribe_local(model: WhisperForConditionalGeneration, audio_16k: np.ndarray) -> str:
|
| 107 |
+
inputs = PROCESSOR(audio_16k, sampling_rate=TARGET_SR, return_tensors="pt")
|
| 108 |
+
feats = inputs.input_features.to(DEVICE)
|
| 109 |
+
if DEVICE == "cuda":
|
| 110 |
+
feats = feats.to(dtype=DTYPE)
|
| 111 |
+
|
| 112 |
+
ids = model.generate(inputs=feats)
|
| 113 |
+
text = PROCESSOR.batch_decode(ids, skip_special_tokens=True)[0]
|
| 114 |
+
return text.strip()
|
| 115 |
+
|
| 116 |
+
|
| 117 |
+
def transcribe_openai(audio_path: str) -> str:
|
| 118 |
+
if not os.getenv("OPENAI_API_KEY"):
|
| 119 |
+
return "OpenAI ASR disabled: OPENAI_API_KEY not set in Space Secrets."
|
| 120 |
+
|
| 121 |
+
prompt = "Chichewa transcription. Malawi names like Lilongwe, Blantyre, Zomba. Keep local names as spoken."
|
| 122 |
+
|
| 123 |
+
with open(audio_path, "rb") as f:
|
| 124 |
+
resp = OPENAI_CLIENT.audio.transcriptions.create(
|
| 125 |
+
file=f,
|
| 126 |
+
model=OPENAI_MODEL,
|
| 127 |
+
language=OPENAI_LANGUAGE, # set to None for auto-detect
|
| 128 |
+
prompt=prompt,
|
| 129 |
+
temperature=0.0,
|
| 130 |
+
response_format="json",
|
| 131 |
+
)
|
| 132 |
+
|
| 133 |
+
return (resp.text or "").strip()
|
| 134 |
+
|
| 135 |
+
|
| 136 |
+
def transcribe_all(audio_path: Optional[str]) -> Tuple[str, str, str, str]:
|
| 137 |
+
"""
|
| 138 |
+
Returns:
|
| 139 |
+
status, base_text, finetuned_text, openai_text
|
| 140 |
+
"""
|
| 141 |
+
if not audio_path:
|
| 142 |
+
return "Please record or upload an audio clip.", "", "", ""
|
| 143 |
+
|
| 144 |
+
# Load audio once
|
| 145 |
+
try:
|
| 146 |
+
y, sr, dur = load_audio(audio_path)
|
| 147 |
+
except Exception as e:
|
| 148 |
+
return f"❌ Failed to load audio: {e}", "", "", ""
|
| 149 |
+
|
| 150 |
+
if dur > MAX_SECONDS:
|
| 151 |
+
return f"⚠️ Audio is {dur:.1f}s. Please keep clips ≤ {MAX_SECONDS:.0f}s.", "", "", ""
|
| 152 |
+
|
| 153 |
+
status = []
|
| 154 |
+
|
| 155 |
+
# Base (local)
|
| 156 |
+
t0 = time.time()
|
| 157 |
+
try:
|
| 158 |
+
base_text = transcribe_local(MODEL_BASE, y)
|
| 159 |
+
status.append(f"✅ Open Source (base) {time.time()-t0:.2f}s")
|
| 160 |
+
except Exception as e:
|
| 161 |
+
base_text = f"[ERROR] Base failed: {e}"
|
| 162 |
+
status.append("❌ Base failed")
|
| 163 |
+
|
| 164 |
+
# Fine-tuned (local)
|
| 165 |
+
t1 = time.time()
|
| 166 |
+
try:
|
| 167 |
+
ft_text = transcribe_local(MODEL_FT, y)
|
| 168 |
+
status.append(f"✅ Fine-tuned {time.time()-t1:.2f}s")
|
| 169 |
+
except Exception as e:
|
| 170 |
+
ft_text = f"[ERROR] Fine-tuned failed: {e}"
|
| 171 |
+
status.append("❌ Fine-tuned failed")
|
| 172 |
+
|
| 173 |
+
# OpenAI (commercial)
|
| 174 |
+
t2 = time.time()
|
| 175 |
+
try:
|
| 176 |
+
openai_text = transcribe_openai(audio_path)
|
| 177 |
+
status.append(f"✅ OpenAI ({OPENAI_MODEL}) {time.time()-t2:.2f}s")
|
| 178 |
+
except Exception as e:
|
| 179 |
+
openai_text = f"[ERROR] OpenAI failed: {e}"
|
| 180 |
+
status.append("❌ OpenAI failed")
|
| 181 |
+
|
| 182 |
+
return "\n".join(status), base_text, ft_text, openai_text
|
| 183 |
+
|
| 184 |
+
|
| 185 |
+
# -----------------------------
|
| 186 |
+
# Warm-up (local models only)
|
| 187 |
+
# -----------------------------
|
| 188 |
+
def warmup():
|
| 189 |
+
try:
|
| 190 |
+
dummy = np.zeros(int(TARGET_SR * 1.0), dtype=np.float32)
|
| 191 |
+
_ = transcribe_local(MODEL_BASE, dummy)
|
| 192 |
+
_ = transcribe_local(MODEL_FT, dummy)
|
| 193 |
+
print("Warm-up complete.", flush=True)
|
| 194 |
+
except Exception as e:
|
| 195 |
+
print(f"Warm-up skipped/failed: {e}", flush=True)
|
| 196 |
+
|
| 197 |
+
|
| 198 |
+
warmup()
|
| 199 |
+
|
| 200 |
+
|
| 201 |
+
# -----------------------------
|
| 202 |
+
# UI
|
| 203 |
+
# -----------------------------
|
| 204 |
+
with gr.Blocks(theme="grass", title="Chichewa Speech2Text") as demo:
|
| 205 |
+
gr.Markdown(LOGO_HTML)
|
| 206 |
+
gr.Markdown(TITLE_HTML)
|
| 207 |
+
gr.Markdown(HIGHLIGHT_TEXT)
|
| 208 |
+
gr.Markdown(DESCRIPTION_HTML)
|
| 209 |
+
|
| 210 |
+
audio_in = gr.Audio(
|
| 211 |
+
sources=["microphone", "upload"],
|
| 212 |
+
type="filepath",
|
| 213 |
+
label="Audio Input (Record or Upload)",
|
| 214 |
+
)
|
| 215 |
+
|
| 216 |
+
run_btn = gr.Button("Transcribe & Compare", variant="primary")
|
| 217 |
+
|
| 218 |
+
status_out = gr.Textbox(label="Status / timing", lines=3)
|
| 219 |
+
|
| 220 |
+
with gr.Row():
|
| 221 |
+
base_out = gr.Textbox(
|
| 222 |
+
label="Open Source ASR Model",
|
| 223 |
+
lines=12,
|
| 224 |
+
)
|
| 225 |
+
ft_out = gr.Textbox(
|
| 226 |
+
label="Open Source Model Fine-Tuned with Custom Chichewa Speech",
|
| 227 |
+
lines=12,
|
| 228 |
+
)
|
| 229 |
+
commercial_out = gr.Textbox(
|
| 230 |
+
label="Frontier Commercial ASR Model (OpenAI)",
|
| 231 |
+
lines=12,
|
| 232 |
+
)
|
| 233 |
+
|
| 234 |
+
run_btn.click(
|
| 235 |
+
fn=transcribe_all,
|
| 236 |
+
inputs=[audio_in],
|
| 237 |
+
outputs=[status_out, base_out, ft_out, commercial_out],
|
| 238 |
+
)
|
| 239 |
+
|
| 240 |
+
gr.Markdown(ARTICLE_HTML)
|
| 241 |
+
|
| 242 |
+
if __name__ == "__main__":
|
| 243 |
+
# Queue helps when multiple people test at once during your seminar
|
| 244 |
+
demo.queue(default_concurrency_limit=2).launch()
|
requirements.txt
ADDED
|
@@ -0,0 +1,84 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
aiofiles==24.1.0
|
| 2 |
+
annotated-doc==0.0.4
|
| 3 |
+
annotated-types==0.7.0
|
| 4 |
+
anyio==4.12.1
|
| 5 |
+
audioread==3.1.0
|
| 6 |
+
brotli==1.2.0
|
| 7 |
+
certifi==2026.1.4
|
| 8 |
+
cffi==2.0.0
|
| 9 |
+
charset-normalizer==3.4.4
|
| 10 |
+
click==8.3.1
|
| 11 |
+
decorator==5.2.1
|
| 12 |
+
distro==1.9.0
|
| 13 |
+
fastapi==0.129.2
|
| 14 |
+
ffmpy==1.0.0
|
| 15 |
+
filelock==3.24.3
|
| 16 |
+
fsspec==2026.2.0
|
| 17 |
+
gradio==6.6.0
|
| 18 |
+
gradio-client==2.1.0
|
| 19 |
+
groovy==0.1.2
|
| 20 |
+
groq==1.0.0
|
| 21 |
+
h11==0.16.0
|
| 22 |
+
hf-xet==1.2.0
|
| 23 |
+
httpcore==1.0.9
|
| 24 |
+
httpx==0.28.1
|
| 25 |
+
huggingface-hub==1.4.1
|
| 26 |
+
idna==3.11
|
| 27 |
+
jinja2==3.1.6
|
| 28 |
+
jiter==0.13.0
|
| 29 |
+
joblib==1.5.3
|
| 30 |
+
lazy-loader==0.4
|
| 31 |
+
librosa==0.11.0
|
| 32 |
+
llvmlite==0.46.0
|
| 33 |
+
markdown-it-py==4.0.0
|
| 34 |
+
markupsafe==3.0.3
|
| 35 |
+
mdurl==0.1.2
|
| 36 |
+
mpmath==1.3.0
|
| 37 |
+
msgpack==1.1.2
|
| 38 |
+
networkx==3.6.1
|
| 39 |
+
numba==0.64.0
|
| 40 |
+
numpy==2.4.2
|
| 41 |
+
openai==2.21.0
|
| 42 |
+
orjson==3.11.7
|
| 43 |
+
packaging==26.0
|
| 44 |
+
pandas==3.0.1
|
| 45 |
+
pillow==12.1.1
|
| 46 |
+
platformdirs==4.9.2
|
| 47 |
+
pooch==1.9.0
|
| 48 |
+
pycparser==3.0
|
| 49 |
+
pydantic==2.12.5
|
| 50 |
+
pydantic-core==2.41.5
|
| 51 |
+
pydub==0.25.1
|
| 52 |
+
pygments==2.19.2
|
| 53 |
+
python-dateutil==2.9.0.post0
|
| 54 |
+
python-multipart==0.0.22
|
| 55 |
+
pytz==2025.2
|
| 56 |
+
pyyaml==6.0.3
|
| 57 |
+
regex==2026.2.19
|
| 58 |
+
requests==2.32.5
|
| 59 |
+
rich==14.3.3
|
| 60 |
+
safehttpx==0.1.7
|
| 61 |
+
safetensors==0.7.0
|
| 62 |
+
scikit-learn==1.8.0
|
| 63 |
+
scipy==1.17.0
|
| 64 |
+
semantic-version==2.10.0
|
| 65 |
+
setuptools==82.0.0
|
| 66 |
+
shellingham==1.5.4
|
| 67 |
+
six==1.17.0
|
| 68 |
+
sniffio==1.3.1
|
| 69 |
+
soundfile==0.13.1
|
| 70 |
+
soxr==1.0.0
|
| 71 |
+
starlette==0.52.1
|
| 72 |
+
sympy==1.14.0
|
| 73 |
+
threadpoolctl==3.6.0
|
| 74 |
+
tokenizers==0.22.2
|
| 75 |
+
tomlkit==0.13.3
|
| 76 |
+
torch==2.10.0
|
| 77 |
+
tqdm==4.67.3
|
| 78 |
+
transformers==5.2.0
|
| 79 |
+
typer==0.24.1
|
| 80 |
+
typer-slim==0.24.0
|
| 81 |
+
typing-extensions==4.15.0
|
| 82 |
+
typing-inspection==0.4.2
|
| 83 |
+
urllib3==2.6.3
|
| 84 |
+
uvicorn==0.41.0
|