04 / RESULT
Transcription
—
LanguageEnglish
Confidence proxy—
Latency—
Real-time factor—
Responsible-use and audio privacy notice
This educational demo can make mistakes. Do not record or upload private, confidential, sensitive, proprietary, medical, legal, safety-critical, surveillance, copyrighted, or personally identifiable audio. Timestamps are approximate and the confidence indicator is only a proxy. Human review is required.
01 / INPUT
Provide audio
●
Record a safe demo clip
Allow microphone access, speak clearly, then stop before transcription.
00:00
Public demonstration sample
Loads a short speech sample bundled with the static app.
02 / MODEL
Configure inference
The public baseline runs live in the browser. The project-exported FP32 ONNX option becomes available after parity validation and model publication. Q8 is retained as an experimental optimization artifact because it did not meet the project quality gate.
06 / EVALUATION
Model evaluation dashboard
Metrics are loaded from static evaluation artifacts generated by the Python scripts. Blank values remain pending; the app never invents portfolio results.
Baseline vs fine-tuned
Run the comparison script to populate this section.
Robustness slices
Noise and accent-slice results are pending.
MODEL DETAILS
Why Whisper Tiny English?
It provides a strong balance between encoder-decoder ASR capability, local LoRA fine-tuning feasibility, ONNX exportability, browser download size, and WebGPU/WASM latency. The English-only checkpoint is deliberately labeled as English rather than claiming multilingual detection.
LIMITATIONS
What this demo is not
This is streaming-style chunked transcription after recording or upload—not a claim of continuous, low-latency production streaming. Confidence is not calibrated, timestamps can drift, browser support varies, and noisy or accented speech can reduce accuracy.
PORTFOLIO VALUE
Applied speech intelligence
The workflow connects to quality meetings, service notes, operator feedback, complaint summaries, maintenance logs, audit notes, and root-cause discussions while preserving responsible data-handling boundaries.