asadiqui/Darija-medical-ASR-benchmark

Benchmarking ASR models (Whisper, Wav2Vec2, SeamlessM4T) on Moroccan Darija medical consultations. # Darija Medical ASR Benchmark This repository contains benchmarking code for three speech recognition models on Moroccan Darija medical consultations. ## Methodology Three speech recognition models were tested on three 15-30 second audio recordings simulating medical consultations: * **Audio 1**: Pure Moroccan Darija symptoms. * **Audio 2**: Mixed Darija and French symptoms. * **Audio 3**: Pure French symptoms. ### Models compared: * **OpenAI Whisper (large-v3)**: `openai/whisper-large-v3` * **Facebook wav2vec2-large-xlsr-53**: Fine-tuned `boumehdi/wav2vec2-large-xlsr-moroccan-darija` * **Meta SeamlessM4T (v2-large)**: `facebook/seamless-m4t-v2-large` ## Transcripts (Reference) ### Transcript 1: Pure Moroccan Darija symptoms > سلام دكتور، هادي يومين وانا مريض. راسي كيضرني بزاف وكرشي حتا هيا. ما قدرتش نعس فاليل وكيجيني رضان. السخانة طالعة ويديا كيفشلو عليا، مابقا عندي جهد لوالو. ### Transcript 2: Moroccan Darija and French mixed symptoms > دكتور، فيا la fièvre et راسي كيضرني بزاف. عندي des problèmes respiratoires والصدر ديالي كيحرقني. Je me sens très fatigué وما بقيتش قادر نتنفس مزيان. ### Transcript 3: Pure French symptoms > Bonjour docteur. J'ai une forte migraine depuis trois jours. J'ai aussi mal à la gorge, une toux sèche persistante et je me sens extrêmement épuisé. La fièvre n'a pas baissé depuis hier matin. ## Setup & Installation 1. Install dependencies: ```bash pip install -r requirements.txt ``` 2. Place your audio files and transcripts in the following structure: - `audio_files/`: Contains `audio1.wav`, `audio2.wav`, `audio3.wav` - `original_transcripts/`: Contains `originalTranscript1.txt`, `originalTranscript2.txt`, `originalTranscript3.txt` - `transcription_output_example/`: (Optional) Contains sample output transcripts for reference. ## Usage 1. **Run Transcription**: Generate transcripts using the three models. ```bash python transcribe.py ``` This will produce output files in the `transcription_output/` folder. 2. **Evaluate Perform …