Upload README.md with huggingface_hub

README.md

@@ -22,11 +22,11 @@ model-index:
       type: automatic-speech-recognition
     metrics:
     - type: wer
-      value: 46.52
-      name: WER
+      value: 26.53
+      name: WER (normalized, word-level)
     - type: cer
-      value: 12.03
-      name: CER
+      value: 11.32
+      name: CER (normalized, word-level)
 ---
 
 # Badini Kurdish ASR (Wav2Vec2-BERT)
@@ -38,17 +38,24 @@ with a CTC head.
 
 ## Results
 
-Evaluated on a held-out test set of **2,038** samples:
+The raw model output (greedy CTC, no post-processing) and the output after the
+project's text-normalization pipeline are reported separately, because word error
+rate is highly sensitive to punctuation, digit spelling, and character variants
+that the normalizer resolves.
 
-| Metric | Score  |
-|--------|--------|
-| WER    | 46.52% |
-| CER    | 12.03% |
+| Evaluation | WER | CER |
+|------------|-----|-----|
+| Raw greedy CTC (no normalization) | 46.5% | 12.0% |
+| Normalized, punctuation stripped (word-level) | **26.5%** | 11.3% |
+| Normalized, punctuation kept | 32.4% | 14.2% |
+
+Normalized results are on a 1,123-sample test set; the raw figure is from the
+training-time evaluation on 2,038 samples.
 
 ## Input requirements
 
 - Audio must be **16 kHz, mono**.
-- Resample any other sample rate to 16 kHz before running inference.
+- Resample any other sample rate to 16 kHz before inference.
 
 ## Usage
 
@@ -70,14 +77,14 @@ print(processor.batch_decode(pred_ids)[0])
 
 ## Important: text normalization
 
-This repository contains **only the acoustic model**. The raw output is plain CTC
-decoding. A separate normalization pipeline — punctuation restoration,
-number-to-word expansion (with Badini-specific remapping), character
-normalization, and word-variant correction — is applied in the downstream
-application and is **not** included here. Output straight from this model will
-not match the fully normalized transcripts.
+This repository contains **only the acoustic model**. Raw output is plain CTC
+decoding. The normalization pipeline that produces the 26.5% WER — punctuation
+handling, number-to-word expansion (Badini-specific), character normalization,
+and word-variant correction — is applied downstream and is **not** included here.
+Output straight from this model will match the "raw" row above, not the normalized
+rows.
 
 ## Intended use
 
-Speech transcription for Badini Kurdish. Given the WER, best used with
-downstream normalization/post-processing rather than as raw final text.
+Speech transcription for Badini Kurdish, best used together with the downstream
+normalization/post-processing described above.