How margotmassonresearch/linguatest works
LinguaTest generates controlled synthetic accent-like speech and evaluates ASR robustness. Its principal workflow selects clean reference sentences, chooses phoneme substitutions from an accent profile, synthesizes variants, transcribes them, and aligns outputs for error analysis. The graph also preserves profile construction, supported recognizers, and evaluation/visualization capabilities. The README and excerpts establish the main orchestration and several internal calls; links involving unsampled modules are included only where the README documents them, and some implementation details remain unverified.
Main entry points: asr_from_name.py Recognizer selection · linguatest.py Pipeline runner · sentence_selection_module.py Clean sentence selection · weighted_alignment.py Weighted alignment
Components 21
Test generation
- Pipeline runner
linguatest.py - Clean sentence selection
sentence_selection_module.py - Phoneme selection
phoneme_selection_module.py - Variant generation
test_utterances_generation_module.py - Speech synthesis
synthesis/azure_synthesis.py
Phonetic processing
- Accent profile builder
synthesis/build_profile.py - Phoneme conversion
phonemes_processing/phonemizer.py - Phoneme preprocessing
phonemes_processing/preprocessing.py
Speech recognition
- Recognizer selection
automatic_speech_recognition/asr_from_name.py - Wav2Vec2 recognizer
automatic_speech_recognition/wav2vec2.py - Wav2Vec2 phonemes
automatic_speech_recognition/wav2vec2phoneme.py - Whisper recognizer
automatic_speech_recognition/whisper.py - SpeechT5 recognizer
automatic_speech_recognition/speecht5.py
Evaluation and analysis
- Robustness assessment
robustness_assessment.py - Dataframe evaluation
evaluation/df_evaluation.py - Weighted alignment
evaluation/weighted_alignment.py - Weighted error scoring
evaluation/weighted_jiwer.py - Phoneme confusion analysis
evaluation/confusion_matrix.py - Analysis visualization
evaluation/visualisation.py
Other components
- Framework user
- Azure TTS
Connections 23
- Framework user → Pipeline runner: runs pipeline
README.md - Pipeline runner → Clean sentence selection: selects sentences
linguatest.py - Pipeline runner → Phoneme selection: selects phonemes
linguatest.py - Pipeline runner → Variant generation: generates variants
linguatest.py - Pipeline runner → Robustness assessment: runs assessment
linguatest.py - Clean sentence selection → Phoneme conversion: phonemizes text
sentence_selection_module.py - Clean sentence selection → Speech synthesis: synthesizes baseline
sentence_selection_module.py - Clean sentence selection → Recognizer selection: chooses recognizer
sentence_selection_module.py - Clean sentence selection → Dataframe evaluation: evaluates baseline
sentence_selection_module.py - Accent profile builder → Phoneme confusion analysis: builds from errors
synthesis/build_profile.py - Variant generation → Speech synthesis: synthesizes variants
- Speech synthesis → Azure TTS: requests speech
synthesis/azure_synthesis.py - Robustness assessment → Recognizer selection: runs recognizer
README.md - Robustness assessment → Weighted alignment: aligns output
README.md - Recognizer selection → Wav2Vec2 recognizer: dispatches to
automatic_speech_recognition/asr_from_name.py - Recognizer selection → Wav2Vec2 phonemes: dispatches to
automatic_speech_recognition/asr_from_name.py - Recognizer selection → Whisper recognizer: dispatches to
automatic_speech_recognition/asr_from_name.py - Recognizer selection → SpeechT5 recognizer: dispatches to
automatic_speech_recognition/asr_from_name.py - Dataframe evaluation → Weighted alignment: aligns transcripts
- Weighted alignment → Phoneme preprocessing: preprocesses phones
evaluation/weighted_alignment.py - Weighted alignment → Weighted error scoring: scores phone errors
evaluation/weighted_alignment.py - Accent profile builder → Wav2Vec2 phonemes: transcribes profile data
synthesis/build_profile.py - Analysis visualization → Phoneme confusion analysis: visualizes analysis