How ashwinhegde19/mendspeech works
MendSpeech is an early-build dictation and speech-robustness project. Its demonstrated workflow loads and damages benchmark audio for comparison; the recognition and scoring modules provide a foundation for evaluating transcripts and confidence. SpeechDamageBench is also a standalone robustness suite. Streaming, triage, editing, post-training, and serving are documented product capabilities but are planned, not represented by source files in the tree. Connections are limited to excerpts, SOURCE INDEX references, and explicit README statements.
Main entry points: audio_lab.py Audio comparison lab · baseline.py Wav2Vec2 baseline · features.py Log-Mel features · audio_damage.py Damage operators
Components 21
Audio processing
- Waveform loader
src/audio/loader.py - Spectrogram analysis
src/audio/stft.py - Log-Mel features
src/audio/features.py
Robustness lab
- Audio comparison lab
app/audio_lab.py - Damage operators
speechdamagebench/speechdamagebench/audio_damage.py - Damage presets
speechdamagebench/speechdamagebench/presets.py - Benchmark manifest
data/benchmark/manifest.csv
Recognition
- Wav2Vec2 baseline
src/asr/baseline.py - CTC decoding
src/asr/ctc_decode.py - Token confidence
src/asr/confidence.py
Evaluation
- Audio pair metrics
src/metrics/audio_metrics.py - Text normalization
src/metrics/normalization.py - WER / CER scoring
src/metrics/wer.py - Protected-content slices
src/metrics/slices.py - Benchmark role audit
src/metrics/roles.py
Planned dictation
- Streaming recognition
- Confidence triage
- Conservative editor
- Editor post-training
- Async serving
Other components
- Engineer / evaluator
Connections 18
- Engineer / evaluator → Audio comparison lab: runs comparison
app/audio_lab.py - Audio comparison lab → Benchmark manifest: loads benchmark
app/audio_lab.py - Audio comparison lab → Waveform loader: loads clips
app/audio_lab.py - Audio comparison lab → Damage operators: applies corruption
app/audio_lab.py - Damage operators → Damage presets: resolves parameters
speechdamagebench/speechdamagebench/__init__.py - Audio comparison lab → Audio pair metrics: measures pairs
app/audio_lab.py - Audio pair metrics → Log-Mel features: computes log-Mel distance
src/metrics/audio_metrics.py - Log-Mel features → Spectrogram analysis: uses transform
src/audio/features.py - Wav2Vec2 baseline → Waveform loader: loads audio
src/asr/baseline.py - Wav2Vec2 baseline → CTC decoding: decodes emissions
- Wav2Vec2 baseline → Token confidence: provides confidence inputs
- Text normalization → WER / CER scoring: normalizes scoring text
src/metrics/wer.py - WER / CER scoring → Protected-content slices: shares alignment convention
src/metrics/slices.py - Engineer / evaluator → Streaming recognition: dictates
README.md - Streaming recognition → Confidence triage: routes by confidence
README.md - Confidence triage → Conservative editor: selects editing
README.md - Conservative editor → Editor post-training: receives bounded training
README.md - Streaming recognition → Async serving: runs through endpoint
README.md