add evaluation script and results

README.md

@@ -8,6 +8,37 @@ tags:
 - robust-speech-event
 datasets:
 - mozilla-foundation/common_voice_7_0
+model-index:
+- name: mpoyraz/wav2vec2-xls-r-300m-cv7-turkish
+  results:
+  - task: 
+       name: Automatic Speech Recognition 
+       type: automatic-speech-recognition
+    dataset:
+       name: Common Voice 7
+       type: mozilla-foundation/common_voice_7_0
+       args: tr
+    metrics:
+      - name: Test WER
+        type: wer
+        value: 20.11
+      - name: Test CER
+        type: cer
+        value: 8.01
+  - task: 
+       name: Automatic Speech Recognition
+       type: automatic-speech-recognition
+    dataset:
+       name: Robust Speech Event - Dev Data
+       type: speech-recognition-community-v2/dev_data
+       args: tr
+    metrics:
+      - name: Test WER
+        type: wer
+        value: 30.87
+      - name: Test CER
+        type: cer
+        value: 10.68
 ---
 
 # wav2vec2-xls-r-300m-cv7-turkish
@@ -47,3 +78,22 @@ The following hypermaters were used for finetuning:
 
 ## Language Model
 N-gram language model is trained on a Turkish Wikipedia articles using KenLM and [ngram-lm-wiki](https://github.com/mpoyraz/ngram-lm-wiki) repo was used to generate arpa LM and convert it into binary format.
+
+## Evaluation Commands
+Please install [unicode_tr](https://pypi.org/project/unicode_tr/) package before running evaluation. It is used for Turkish text processing.
+1. To evaluate on `mozilla-foundation/common_voice_7_0` with split `test`
+```bash
+python eval.py --model_id mpoyraz/wav2vec2-xls-r-300m-cv7-turkish --dataset mozilla-foundation/common_voice_7_0 --config tr --split test
+```
+
+2. To evaluate on `speech-recognition-community-v2/dev_data`
+
+```bash
+python eval.py --model_id mpoyraz/wav2vec2-xls-r-300m-cv7-turkish --dataset speech-recognition-community-v2/dev_data --config tr --split validation --chunk_length_s 5.0 --stride_length_s 1.0
+```
+## Evaluation results:
+
+| Dataset | WER | CER |
+|---|---|---|
+|Common Voice 7 TR test split| 20.11 | 8.01 |
+|Speech Recognition Community dev data| 30.87 | 10.68 |

eval.py

@@ -0,0 +1,151 @@
+#!/usr/bin/env python3
+import argparse
+import re
+from typing import Dict
+
+import torch
+from datasets import Audio, Dataset, load_dataset, load_metric
+
+from transformers import AutoFeatureExtractor, pipeline
+
+from unicode_tr import unicode_tr
+
+def log_results(result: Dataset, args: Dict[str, str]):
+    """DO NOT CHANGE. This function computes and logs the result metrics."""
+
+    log_outputs = args.log_outputs
+    dataset_id = "_".join(args.dataset.split("/") + [args.config, args.split])
+
+    # load metric
+    wer = load_metric("wer")
+    cer = load_metric("cer")
+
+    # compute metrics
+    wer_result = wer.compute(references=result["target"], predictions=result["prediction"])
+    cer_result = cer.compute(references=result["target"], predictions=result["prediction"])
+
+    # print & log results
+    result_str = f"WER: {wer_result}\n" f"CER: {cer_result}"
+    print(result_str)
+
+    with open(f"{dataset_id}_eval_results.txt", "w") as f:
+        f.write(result_str)
+
+    # log all results in text file. Possibly interesting for analysis
+    if log_outputs is not None:
+        pred_file = f"log_{dataset_id}_predictions.txt"
+        target_file = f"log_{dataset_id}_targets.txt"
+
+        with open(pred_file, "w") as p, open(target_file, "w") as t:
+
+            # mapping function to write output
+            def write_to_file(batch, i):
+                p.write(f"{i}" + "\n")
+                p.write(batch["prediction"] + "\n")
+                t.write(f"{i}" + "\n")
+                t.write(batch["target"] + "\n")
+
+            result.map(write_to_file, with_indices=True)
+
+
+def normalize_text(text: str) -> str:
+    """DO ADAPT FOR YOUR USE CASE. this function normalizes the target text."""
+
+    chars_to_remove_regex = '[,?.!\-\;\:"“%”�—…–()]'
+    apostrophes = "[’‘`´ʹʻʼʽʿˈ]"
+    
+    # Lower the text using 'unicode_tr'
+    # Regular lower() does not work well for Turkish Language
+    text_norm = unicode_tr(text).lower()
+    # Unify apostrophes
+    text_norm = re.sub(apostrophes, "'", text_norm)
+    # Remove pre-defined chars
+    text_norm = re.sub(chars_to_remove_regex, "", text_norm)
+    # Remove single quotes
+    text_norm = text_norm.replace(" '", " ")
+    text_norm = text_norm.replace("' ", " ")
+    # Handle hatted characters
+    text_norm = re.sub('[â]', 'a', text_norm)
+    text_norm = re.sub('[î]', 'i', text_norm)
+    text_norm = re.sub('[ô]', 'o', text_norm)
+    text_norm = re.sub('[û]', 'u', text_norm)
+    # Handle alternate characters
+    text_norm = re.sub('[é]', 'e', text_norm)
+    text_norm = re.sub('[ë]', 'e', text_norm)
+    # Remove multiple spaces
+    text_norm = re.sub(r"\s+", " ", text_norm)
+
+    return text_norm
+
+
+def main(args):
+    # load dataset
+    dataset = load_dataset(args.dataset, args.config, split=args.split, use_auth_token=True)
+
+    # for testing: only process the first two examples as a test
+    # dataset = dataset.select(range(10))
+
+    # load processor
+    feature_extractor = AutoFeatureExtractor.from_pretrained(args.model_id)
+    sampling_rate = feature_extractor.sampling_rate
+
+    # resample audio
+    dataset = dataset.cast_column("audio", Audio(sampling_rate=sampling_rate))
+
+    # load eval pipeline
+    if args.device is None:
+        args.device = 0 if torch.cuda.is_available() else -1
+    asr = pipeline("automatic-speech-recognition", model=args.model_id, device=args.device)
+
+    # map function to decode audio
+    def map_to_pred(batch):
+        prediction = asr(
+            batch["audio"]["array"], chunk_length_s=args.chunk_length_s, stride_length_s=args.stride_length_s
+        )
+
+        batch["prediction"] = prediction["text"]
+        batch["target"] = normalize_text(batch["sentence"])
+        return batch
+
+    # run inference on all examples
+    result = dataset.map(map_to_pred, remove_columns=dataset.column_names)
+
+    # compute and log_results
+    # do not change function below
+    log_results(result, args)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--model_id", type=str, required=True, help="Model identifier. Should be loadable with 🤗 Transformers"
+    )
+    parser.add_argument(
+        "--dataset",
+        type=str,
+        required=True,
+        help="Dataset name to evaluate the `model_id`. Should be loadable with 🤗 Datasets",
+    )
+    parser.add_argument(
+        "--config", type=str, required=True, help="Config of the dataset. *E.g.* `'en'`  for Common Voice"
+    )
+    parser.add_argument("--split", type=str, required=True, help="Split of the dataset. *E.g.* `'test'`")
+    parser.add_argument(
+        "--chunk_length_s", type=float, default=None, help="Chunk length in seconds. Defaults to 5 seconds."
+    )
+    parser.add_argument(
+        "--stride_length_s", type=float, default=None, help="Stride of the audio chunks. Defaults to 1 second."
+    )
+    parser.add_argument(
+        "--log_outputs", action="store_true", help="If defined, write outputs to log file for analysis."
+    )
+    parser.add_argument(
+        "--device",
+        type=int,
+        default=None,
+        help="The device to run the pipeline on. -1 for CPU (default), 0 for the first GPU and so on.",
+    )
+    args = parser.parse_args()
+
+    main(args)