---
language:
- tw
- ak
license: cc-by-sa-4.0
tags:
- automatic-speech-recognition
- whisper
- akan
- twi
- ghanaian-speech-lab
- serendepify-gsl
pipeline_tag: automatic-speech-recognition
base_model: teckedd/whisper-small-waxal-round2-specaug-v1
library_name: transformers
---

# serendepify-gsl-asr-ak-waxal-gnlp-whisper-small-broad-lowlr-freezeenc-fullft-v0.3

This is the third Ghanaian Speech Lab ASR artifact: a broader Waxal+GhanaNLP
low-learning-rate full fine-tuning pass with the Whisper encoder frozen.

It is a real training run, not a smoke-test checkpoint, but it is still a
review candidate until the held-out evaluation and failure taxonomy are complete.

License is set conservatively because this pass includes Waxal-derived data.

- Training rows: 12000
- Dev rows: 512
- Max steps: 800
- Method: full fine-tuning from `teckedd/whisper-small-waxal-round2-specaug-v1`
- Dataset mix: broad Waxal + GhanaNLP from the v0.1 sanitized manifest
- Frozen encoder: True
- Stop on first regression: True

Baseline metrics:

```json
{
  "baseline_loss": 5.13262939453125,
  "baseline_model_preparation_time": 0.0046,
  "baseline_wer": 0.4543915741959752,
  "baseline_runtime": 238.4489,
  "baseline_samples_per_second": 2.147,
  "baseline_steps_per_second": 0.537
}
```

Baseline by corpus:

```json
{
  "waxal": {
    "baseline_waxal_loss": 3.3425865173339844,
    "baseline_waxal_model_preparation_time": 0.0046,
    "baseline_waxal_wer": 0.3454323454323454,
    "baseline_waxal_runtime": 158.7616,
    "baseline_waxal_samples_per_second": 1.612,
    "baseline_waxal_steps_per_second": 0.403
  },
  "gnlp": {
    "baseline_gnlp_loss": 6.922672271728516,
    "baseline_gnlp_model_preparation_time": 0.0046,
    "baseline_gnlp_wer": 1.0584615384615386,
    "baseline_gnlp_runtime": 76.1024,
    "baseline_gnlp_samples_per_second": 3.364,
    "baseline_gnlp_steps_per_second": 0.841
  }
}
```

Final metrics:

```json
{
  "final_loss": 1.140019416809082,
  "final_model_preparation_time": 0.0046,
  "final_wer": 0.451758510438217,
  "final_runtime": 223.9252,
  "final_samples_per_second": 2.286,
  "final_steps_per_second": 0.572,
  "epoch": 0.5333333333333333
}
```

Final by corpus:

```json
{
  "waxal": {
    "final_waxal_loss": 0.498562753200531,
    "final_waxal_model_preparation_time": 0.0046,
    "final_waxal_wer": 0.3584193584193584,
    "final_waxal_runtime": 154.7631,
    "final_waxal_samples_per_second": 1.654,
    "final_waxal_steps_per_second": 0.414,
    "epoch": 0.5333333333333333
  },
  "gnlp": {
    "final_gnlp_loss": 1.7814760208129883,
    "final_gnlp_model_preparation_time": 0.0046,
    "final_gnlp_wer": 0.9692307692307692,
    "final_gnlp_runtime": 71.1072,
    "final_gnlp_samples_per_second": 3.6,
    "final_gnlp_steps_per_second": 0.9,
    "epoch": 0.5333333333333333
  }
}
```