"""OpenPA pronunciation assessment demo."""
from gpu_backend import run_assessment, prepare_artifacts, prepare_request # Import ZeroGPU first.
import html
import json
import logging
import os
from pathlib import Path
from threading import Lock
import gradio as gr
from stage_progress import stage_markup, stream_work
from onnx_runtime import load_audio
ROOT = Path(__file__).resolve().parent
SAMPLE = str(ROOT / "examples/sample.wav")
REFERENCE = "I always thought he was a good person"
EXAMPLES = {item["name"]: item for item in json.loads((ROOT / "examples/catalog.json").read_text())}
MODEL_CACHE = os.getenv("HF_HOME", "/tmp/openpa-hf")
LOCK = Lock()
LABELS = [("total", "Overall"), ("accuracy", "Accuracy"),
("fluency", "Fluency"), ("prosodic", "Prosody")]
def cards(scores=None):
items = []
for key, label in LABELS:
value = "—" if scores is None else f"{scores[key]:.2f}"
items.append(f'
{label}'
f'{value}/ 10
')
return '' + "".join(items) + "
"
def clear_results(mode="Reference-text"):
return cards(), "", [], "Upload a clip or try the example, then assess your pronunciation.", stage_markup(mode)
def set_mode(mode):
return gr.update(visible=mode == "Reference-text"), *clear_results()[:4], stage_markup(mode)
def reset_example(example_name="Scripted speech", mode="Reference-text"):
item = EXAMPLES[example_name]
return str(ROOT / "examples" / item["file"]), gr.update(
value=item["text"], visible=mode == "Reference-text"
), *clear_results(mode)
def assess(audio, mode, text):
if not audio:
raise gr.Error("Please upload an audio clip or load the example.")
if mode not in {"Reference-text", "Reference-free"}:
raise gr.Error("Please select an assessment mode.")
if mode == "Reference-text" and not (text or "").strip():
raise gr.Error("Please enter a reference text, or select Reference-free.")
def work(notify):
notify("audio")
waveform, rate = load_audio(audio)
duration[0] = len(waveform) / rate
with LOCK:
notify("model")
model_dir, asr_dir = prepare_artifacts(mode == "Reference-free")
processor, features = prepare_request(waveform, rate, text, model_dir, asr_dir)
return run_assessment(waveform, rate, text if mode == "Reference-text" else None,
model_dir, asr_dir, processor, features, on_stage=notify)
duration = [0.0]
active = "audio"
yield cards(), "", [], "Assessment in progress…", stage_markup(mode, active)
for kind, payload in stream_work(work):
if kind == "stage":
active = payload
yield gr.skip(), gr.skip(), gr.skip(), "Assessment in progress…", stage_markup(mode, active)
elif kind == "error":
yield cards(), "", [], "Assessment stopped. Please try again.", stage_markup(mode, active, failed=True)
if isinstance(payload, (ValueError, gr.Error)):
raise gr.Error(str(payload)) from payload
logging.error("Assessment failed", exc_info=(type(payload), payload, payload.__traceback__))
raise gr.Error("Assessment could not finish. Please try again in a moment.") from payload
else:
result = payload
rows = [[word["word"], round(word["accuracy"], 2), round(word["total"], 2)] for word in result["words"]]
caption = ("Recognized transcript" if mode == "Reference-free" else "Reference transcript")
transcript = f'{caption}{html.escape(result["transcript"])}
'
yield cards(result["utterance"]), transcript, rows, f"Assessment complete · {mode} · {duration[0]:.1f}s audio", stage_markup(mode, complete=True)
def build_demo():
with gr.Blocks(title="OpenPA · Pronunciation assessment", delete_cache=(3600, 3600)) as demo:
gr.HTML('')
with gr.Row(elem_id="workspace"):
with gr.Column(scale=5, min_width=300, elem_classes="panel"):
gr.HTML('01
Your voice
')
example_choice = gr.Dropdown(list(EXAMPLES), value="Scripted speech",
label="Example recording", interactive=True)
audio = gr.Audio(value=SAMPLE, sources=["upload", "microphone"], type="filepath",
label="Audio clip", interactive=True)
gr.Markdown("English audio · up to 30 seconds", elem_classes="hint")
mode = gr.Radio(["Reference-text", "Reference-free"], value="Reference-text",
label="Assessment mode", interactive=True)
text = gr.Textbox(value=REFERENCE, label="Reference text", lines=3,
placeholder="What did you intend to say?", interactive=True)
gr.Markdown("Reference-free uses WhisperX large-v3 to transcribe your audio before scoring.", elem_classes="hint")
with gr.Row():
submit = gr.Button("Scoring →", variant="primary", elem_id="assess")
example = gr.Button("Reload example", variant="secondary")
with gr.Column(scale=6, min_width=300, elem_classes="panel"):
gr.HTML('02
Your assessment
')
stages = gr.HTML(stage_markup("Reference-text"), elem_id="assessment-stages")
scores = gr.HTML(cards())
transcript = gr.HTML("")
words = gr.Dataframe(headers=["Word", "Accuracy / 10", "Overall / 10"],
datatype=["str", "number", "number"], value=[],
interactive=False, label="Word-level scores", elem_id="word-scores")
status = gr.Markdown(clear_results()[3], elem_classes="hint", elem_id="status")
gr.HTML('')
outputs = [scores, transcript, words, status, stages]
submit.click(assess, [audio, mode, text], outputs, api_name=False,
concurrency_limit=1, concurrency_id="assessment", show_progress="hidden")
mode.change(set_mode, mode, [text, *outputs], queue=False, api_name=False)
audio.change(clear_results, inputs=mode, outputs=outputs, queue=False, api_name=False)
text.input(clear_results, inputs=mode, outputs=outputs, queue=False, api_name=False)
example_choice.change(reset_example, inputs=[example_choice, mode],
outputs=[audio, text, *outputs], queue=False, api_name=False)
example.click(reset_example, inputs=[example_choice, mode],
outputs=[audio, text, *outputs], queue=False, api_name=False)
return demo.queue(max_size=8, default_concurrency_limit=1)
if __name__ == "__main__":
build_demo().launch(
server_name="0.0.0.0", server_port=int(os.getenv("PORT", "7860")),
css=(ROOT / "style.css").read_text(), theme=gr.themes.Base(),
js="() => { document.documentElement.classList.add('dark'); document.body.classList.add('dark'); document.querySelector('.gradio-container').classList.add('dark'); }",
footer_links=[], allowed_paths=[str(ROOT / "examples")],
blocked_paths=[MODEL_CACHE, os.getenv("OPENPA_MODEL_DIR", MODEL_CACHE)],
max_file_size="20mb",
)