kid-whisper-base-en-myst-ours
Fine-tuned openai/whisper-base.en on the MyST children's speech corpus,
for research into compression (quantization + pruning) of children's ASR
models at small scale.
Model Details
- Base model:
openai/whisper-base.en(74M parameters) - Fine-tuning data: MyST corpus, filtered train/dev splits (57,687 train utterances / 9,017 dev utterances)
- Filtering pipeline: custom preprocessing (missing-transcript removal, reference-ASR quality filtering, non-speech-tag handling, short-utterance removal, duration capping for train/dev). See [link to preprocess_myst.py / thesis Chapter 2 after published] for full filter specification.
- Training framework: HuggingFace
Seq2SeqTrainer - Precision: fp16 mixed-precision training
Training Hyperparameters
| Parameter | Value |
|---|---|
| Learning rate | 1e-5 |
| Per-device batch size | 8 |
| Gradient accumulation | 8 (effective batch 64) |
| Warmup steps | 500 |
| Max steps | 8000 (all 8000 completed) |
| Early stopping | patience=5, metric=WER on dev |
| Best checkpoint | step 4500 |
Note on batch configuration. The companion tiny.en model was trained
with per-device batch 16 × gradient accumulation 4. At base.en scale that
configuration exhausts the 8GB VRAM of the training GPU (RTX 4060), so the
micro-batch was halved and accumulation doubled. The effective batch size
is 64 in both cases, keeping the optimization setup equivalent and the
two models directly comparable.
Evaluation
Training-time dev WER (early-stopping metric only): 10.65%
Measured on individual (non-concatenated) data/filtered/dev utterances,
used solely for checkpoint selection during training. This is not the
model's reportable evaluation number.
Training ran the full 8000 steps. Dev WER reached its minimum at step 4500 and did not improve across the eight subsequent evaluations, confirming the plateau rather than stopping early on it:
| Step | 3000 | 4000 | 4500 | 5000 | 5500 | 6000 | 6500 | 7000 | 7500 | 8000 |
|---|---|---|---|---|---|---|---|---|---|---|
| Dev WER % | 10.92 | 10.70 | 10.65 | 10.80 | 10.84 | 10.91 | 10.97 | 10.97 | 10.90 | 10.89 |
Reportable baseline WER: 11.91%
Measured via HuggingFace ASR pipeline with chunk_length_s=30, beam search
(beam=5), batch=4, against the full 3,972-chunk held-out MyST concatenated
test set (data/concatenated/test) — the identical protocol used to
evaluate every other checkpoint in this research.
| Model | Parameters | FP16 Baseline WER |
|---|---|---|
| kid-whisper-tiny-en-myst-ours | 39M | 14.53% |
| kid-whisper-base-en-myst-ours (this model) | 74M | 11.91% |
| kid-whisper-small-en-myst (Attia et al.) | 244M | 9.16% |
| kid-whisper-small-myst, multilingual (Attia et al.) | 244M | 9.91% |
| kid-whisper-medium-en-myst (Attia et al.) | 769M | 8.94% |
All five numbers were measured under the identical evaluation protocol above, giving a clean, monotonically-improving capacity scaling curve (39M → 74M → 244M → 769M) under a single consistent methodology — with no discontinuity between the self-trained checkpoints (tiny-en, base-en) and the externally-trained ones (small-en, small-multilingual, medium-en).
Relationship to Other KID-Whisper Checkpoints
This checkpoint is independently trained and is not derived from or affiliated with:
aadel4/kid-whisper-small-en-myst/aadel4/kid-whisper-medium-en-myst(Attia et al., fine-tuned Small/Medium Whisper on MyST)SatwikDutta/kid-whisper-tiny-en-myst(Dutta et al., fine-tuned tiny.en — a different model size — on a different MyST filtering/split scheme)
No publicly available, ungated, MyST-fine-tuned base.en checkpoint with a
documented filtering scheme matching this study's pipeline was available at
time of training. This checkpoint was therefore trained specifically to
extend a compression study (quantization + pruning) to base-scale Whisper
using the same data filtering and evaluation pipeline already used for
Small/Medium-scale checkpoints in that study, avoiding a training-data
confound that would have resulted from mixing in an externally-trained
checkpoint with a different, undocumented data version.
It is the direct counterpart of kid-whisper-tiny-en-myst-ours, trained on
identical data with an identical protocol and effective batch size; the two
differ only in base model scale.
Intended Use
Research use for studying model compression (post-training quantization, magnitude/importance-based pruning) on children's speech ASR at small model scale. Not evaluated or intended for production deployment.
License
MIT, matching the base openai/whisper-base.en license.