--- language: - adj license: apache-2.0 library_name: transformers pipeline_tag: automatic-speech-recognition tags: - adja - asr - speech-recognition - low-resource - wav2vec2 - xls-r - cs-thesis-may-2026 datasets: - JosueG/adja-speech-asr-tts base_model: facebook/wav2vec2-xls-r-300m --- # Wav2Vec2 XLS-R Adja ASR C4v2 This is the thesis-final deployable CTC ASR model for Adja speech. It is the main public ASR weight location used by the May 2026 CS thesis release. ## Thesis Role This model is the primary Adja speech-to-text artifact. It is used as an ASR baseline, an endpoint model in the S2TT pipeline, and one of the reverse-WER judges for TTS evaluation. ## Model And Data - **Task:** automatic speech recognition - **Base model:** `facebook/wav2vec2-xls-r-300m` - **Training data:** Orpheus Adja speech lineage, public canonical dataset `JosueG/adja-speech-asr-tts` - **Input audio:** use 16 kHz audio at inference time unless your pipeline handles resampling explicitly - **Release repo:** `FrejusGdm/cs-thesis-may-2026` ## Headline Result The C4v2 acoustic model reports **25.05% CER** and **72.39% WER** with greedy decoding in the project registry. With a character language model and Optuna shallow-fusion tuning, the best reported ASR setting reaches **22.67% normalized CER** and **70.76% normalized WER**. See: - `results/adja-nmt/D4_ctc_lm/conclusion.md` - `docs/source-repos/adja-nmt/experiment-registry.md` ## Limitations - The public model repo is the acoustic model surface; some LM decoding experiments live in the result reports rather than inside this model repo. - Deployed endpoint behavior can differ from offline evaluation because of tokenizer reconstruction, chunking, and resampling. - CER is more informative than WER for this release because Adja word-boundary and orthographic normalization choices are still evolving. ## Citation If you use this model, cite: Josue Godeme. 2026. *CS Thesis May 2026: French-Adja MT and Adja Speech Experiments*. https://github.com/FrejusGdm/cs-thesis-may-2026