Fine-tuning an Automatic Speech Recognition model for a Canadian Indigenous counselling program
Bibliographic record
Abstract
Automatic Speech Recognition (ASR) systems have seen a marked improvement in performance after adopting the End-To-End (E2E) approach. However, the performance of ASR models is still largely dependent on the quantity and quality of data they are trained on. Popular ASR systems today are trained on thousands of hours of data but consistently fail to maintain their performance when exposed to outlier accents, vocal pitches, and demographic speech. While ASR systems have greatly improved in recent years, the biggest hurdle remains the lack of niche speech data. Most available speech data fall into a few voice types and are not representative of the average ASR system user. The majority of machine learning projects require large amounts of data. However, the adaptability of E2E ASR models allows them to be fine-tuned to outlier speech using small amounts of representative data. The project presented in this thesis aimed to fine-tune an ASR model for use in an Indigenous Counselling Program. An open-source ASR system called Mozilla DeepSpeech was used to train the models used for the project. Representative speech data was gathered from audiobooks and organized into speech corpora to fine-tune DeepSpeech’s pre-trained models. DeepSpeech’s live transcription software was also implemented on the Unity Development Engine to ensure compatibility with the Indigenous Counselling Program. Three models were trained using the new speech data. Two models were trained using pitch frequency-specific data. One general model was trained using all the new speech data. The results showed a minimum average relative WER or WER improvement of 8.90% for all the models, on the dataset they were trained on. Furthermore, the general model showed little to no improvement in performance over the pitch specific models when tested on their trained datasets. This demonstrated the importance of using representative speech data in ASR model training. Overall, the models showed a marked improvement in performance when trained on the intended user accent and voice type.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.000 | 0.000 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.001 | 0.000 |
| Science and technology studies | 0.001 | 0.000 |
| Scholarly communication | 0.000 | 0.001 |
| Open science | 0.001 | 0.000 |
| Research integrity | 0.000 | 0.000 |
| Insufficient payload (model declined to judge) | 0.000 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one teacher head, not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".