Add model card with Japanese FLEURS evaluation and source commit

README.md

@@ -0,0 +1,36 @@
+---
+language: ja
+tags:
+  - automatic-speech-recognition
+  - japanese
+  - lambda-audio
+datasets:
+  - google/fleurs
+metrics:
+  - cer
+---
+
+# LambdaAudio v6 (350M)
+
+This is a Japanese automatic speech recognition model trained in the LambdaAudio pipeline. The model combines `KeisukeMiyamoto/zipformer-cpt` as its encoder with `KeisukeMiyamoto/lfm2.5-230m-cpt` as its decoder.
+
+## Evaluation
+
+The `model.ckpt` checkpoint achieved **6.05% character error rate (CER)** on the Japanese (`ja_jp`) test subset of [Google FLEURS](https://huggingface.co/datasets/google/fleurs).
+
+- CER: **6.05%**
+- Elapsed time: **140.89 seconds**
+
+CER was calculated over the evaluation corpus after text normalization, punctuation removal, and whitespace removal. Samples without usable text and samples containing predominantly non-Japanese characters were skipped, following the LambdaAudio `src.cer` evaluation code.
+
+## Code Version
+
+LambdaAudio source commit: [`28ba788c5891572c71d93d26b5a2fee52a3297bc`](https://github.com/KeisukeMiyamoto1324/lambda-audio/tree/28ba788c5891572c71d93d26b5a2fee52a3297bc).
+
+## Files
+
+- `model.ckpt`: checkpoint used for the reported evaluation.
+- `last.ckpt`: latest saved training checkpoint.
+- `training_config.json`: training configuration.
+- `logs/`: training metrics.
+- `inference_history.csv`: inference history from training.