diff --git a/README.md b/README.md index f0111c3..eeb316a 100644 --- a/README.md +++ b/README.md @@ -2,8 +2,6 @@ language: en datasets: - common_voice -- librispeech_asr -- timit_asr metrics: - wer - cer @@ -26,15 +24,15 @@ model-index: metrics: - name: Test WER type: wer - value: 19.76 + value: 18.98 - name: Test CER type: cer - value: 8.60 + value: 8.29 --- # Wav2Vec2-Large-XLSR-53-English -Fine-tuned [facebook/wav2vec2-large-xlsr-53](https://huggingface.co/facebook/wav2vec2-large-xlsr-53) on English using the [Common Voice](https://huggingface.co/datasets/common_voice), [LibriSpeech](https://huggingface.co/datasets/librispeech_asr) and [TIMIT](https://huggingface.co/datasets/timit_asr),. +Fine-tuned [facebook/wav2vec2-large-xlsr-53](https://huggingface.co/facebook/wav2vec2-large-xlsr-53) on English using the [Common Voice](https://huggingface.co/datasets/common_voice). When using this model, make sure that your speech input is sampled at 16kHz. The script used for training can be found here: https://github.com/jonatasgrosman/wav2vec2-sprint @@ -83,16 +81,16 @@ for i, predicted_sentence in enumerate(predicted_sentences): | Reference | Prediction | | ------------- | ------------- | -| "SHE'LL BE ALL RIGHT." | SHE'D BE ALRIGHT | +| "SHE'LL BE ALL RIGHT." | SHE'LL BE ALL RIGHT | | SIX | SIX | -| "ALL'S WELL THAT ENDS WELL." | ALL IS WELL THAT ENDS WELL | +| "ALL'S WELL THAT ENDS WELL." | ALL AS WELL THAT ENDS WELL | | DO YOU MEAN IT? | DO YOU MEAN IT | | THE NEW PATCH IS LESS INVASIVE THAN THE OLD ONE, BUT STILL CAUSES REGRESSIONS. | THE NEW PATCH IS LESS INVASIVE THAN THE OLD ONE BUT STILL CAUSES REGRESSION | -| HOW IS MOZILLA GOING TO HANDLE AMBIGUITIES LIKE QUEUE AND CUE? | HOW IS MUSILA GOING TO HANDLE ANB HOOTIES LIKE QU AND QU | -| "I GUESS YOU MUST THINK I'M KINDA BATTY." | RISIONAS INCI IN TE BACTY | +| HOW IS MOZILLA GOING TO HANDLE AMBIGUITIES LIKE QUEUE AND CUE? | HOW IS MOSLILLAR GOING TO HANDLE ANDBEWOOTH HIS LIKE Q AND Q | +| "I GUESS YOU MUST THINK I'M KINDA BATTY." | RUSTIAN WASTIN PAN ONTE BATTLY | | NO ONE NEAR THE REMOTE MACHINE YOU COULD RING? | NO ONE NEAR THE REMOTE MACHINE YOU COULD RING | -| SAUCE FOR THE GOOSE IS SAUCE FOR THE GANDER. | SAUCE FOR THE GUISE IS SAUCE FOR THE GONDER | -| GROVES STARTED WRITING SONGS WHEN SHE WAS FOUR YEARS OLD. | GRAFS STARTED WRITING SOUNDS WHEN SHE WAS FOUR YEARS OLD | +| SAUCE FOR THE GOOSE IS SAUCE FOR THE GANDER. | SAUCE FOR THE GUICE IS SAUCE FOR THE GONDER | +| GROVES STARTED WRITING SONGS WHEN SHE WAS FOUR YEARS OLD. | GRAFS STARTED WRITING SONGS WHEN SHE WAS FOUR YEARS OLD | ## Evaluation @@ -117,11 +115,6 @@ CHARS_TO_IGNORE = [",", "?", "¿", ".", "!", "¡", ";", "；", ":", '""', "%", ' test_dataset = load_dataset("common_voice", LANG_ID, split="test") -# uncomment the following lines to eval using other datasets -# test_dataset = load_dataset("librispeech_asr", "clean", split="test") -# test_dataset = load_dataset("librispeech_asr", "other", split="test") -# test_dataset = load_dataset("timit_asr", split="test") - wer = load_metric("wer.py") # https://github.com/jonatasgrosman/wav2vec2-sprint/blob/main/wer.py cer = load_metric("cer.py") # https://github.com/jonatasgrosman/wav2vec2-sprint/blob/main/cer.py @@ -136,9 +129,9 @@ model.to(DEVICE) def speech_file_to_array_fn(batch): with warnings.catch_warnings(): warnings.simplefilter("ignore") - speech_array, sampling_rate = librosa.load(batch["file"] if "file" in batch else batch["path"], sr=16_000) + speech_array, sampling_rate = librosa.load(batch["path"], sr=16_000) batch["speech"] = speech_array - batch["sentence"] = re.sub(chars_to_ignore_regex, "", batch["text"] if "text" in batch else batch["sentence"]).upper() + batch["sentence"] = re.sub(chars_to_ignore_regex, "", batch["sentence"]).upper() return batch test_dataset = test_dataset.map(speech_file_to_array_fn) @@ -166,76 +159,18 @@ print(f"CER: {cer.compute(predictions=predictions, references=references, chunk_ **Test Result**: -In the table below I report the Word Error Rate (WER) and the Character Error Rate (CER) of the model. I ran the evaluation script described above on other models as well (on 2021-05-20). Note that the table below may show different results from those already reported, this may have been caused due to some specificity of the other evaluation scripts used. Initially, I've tested the model only using the Common Voice dataset. Later I've also tested the model using the LibriSpeech and TIMIT datasets, which are better-behaved datasets than the Common Voice, containing only examples in US English extracted from audiobooks. - ---- - -**Common Voice** +In the table below I report the Word Error Rate (WER) and the Character Error Rate (CER) of the model. I ran the evaluation script described above on other models as well (on 2021-06-17). Note that the table below may show different results from those already reported, this may have been caused due to some specificity of the other evaluation scripts used. | Model | WER | CER | | ------------- | ------------- | ------------- | -| jonatasgrosman/wav2vec2-large-xlsr-53-english | **19.76%** | **8.60%** | -| jonatasgrosman/wav2vec2-large-english | 21.16% | 9.53% | +| jonatasgrosman/wav2vec2-large-xlsr-53-english | **18.98%** | **8.29%** | +| jonatasgrosman/wav2vec2-large-english | 21.53% | 9.66% | | facebook/wav2vec2-large-960h-lv60-self | 22.03% | 10.39% | | facebook/wav2vec2-large-960h-lv60 | 23.97% | 11.14% | +| boris/xlsr-en-punctuation | 29.10% | 10.75% | | facebook/wav2vec2-large-960h | 32.79% | 16.03% | -| boris/xlsr-en-punctuation | 34.81% | 15.51% | | facebook/wav2vec2-base-960h | 39.86% | 19.89% | | facebook/wav2vec2-base-100h | 51.06% | 25.06% | | elgeish/wav2vec2-large-lv60-timit-asr | 59.96% | 34.28% | | facebook/wav2vec2-base-10k-voxpopuli-ft-en | 66.41% | 36.76% | | elgeish/wav2vec2-base-timit-asr | 68.78% | 36.81% | - ---- - -**LibriSpeech (clean)** - -| Model | WER | CER | -| ------------- | ------------- | ------------- | -| facebook/wav2vec2-large-960h-lv60-self | **1.86%** | **0.54%** | -| facebook/wav2vec2-large-960h-lv60 | 2.15% | 0.61% | -| facebook/wav2vec2-large-960h | 2.82% | 0.84% | -| facebook/wav2vec2-base-960h | 3.44% | 1.06% | -| jonatasgrosman/wav2vec2-large-xlsr-53-english | 4.16% | 1.28% | -| facebook/wav2vec2-base-100h | 6.26% | 2.00% | -| jonatasgrosman/wav2vec2-large-english | 8.00% | 2.55% | -| elgeish/wav2vec2-large-lv60-timit-asr | 15.53% | 4.93% | -| boris/xlsr-en-punctuation | 19.28% | 6.45% | -| elgeish/wav2vec2-base-timit-asr | 29.19% | 8.38% | -| facebook/wav2vec2-base-10k-voxpopuli-ft-en | 31.82% | 12.41% | - ---- - -**LibriSpeech (other)** - -| Model | WER | CER | -| ------------- | ------------- | ------------- | -| facebook/wav2vec2-large-960h-lv60-self | **3.89%** | **1.40%** | -| facebook/wav2vec2-large-960h-lv60 | 4.45% | 1.56% | -| facebook/wav2vec2-large-960h | 6.49% | 2.52% | -| jonatasgrosman/wav2vec2-large-xlsr-53-english | 8.82% | 3.42% | -| facebook/wav2vec2-base-960h | 8.90% | 3.55% | -| jonatasgrosman/wav2vec2-large-english | 13.62% | 5.24% | -| facebook/wav2vec2-base-100h | 13.97% | 5.51% | -| boris/xlsr-en-punctuation | 26.40% | 10.11% | -| elgeish/wav2vec2-large-lv60-timit-asr | 28.39% | 12.08% | -| elgeish/wav2vec2-base-timit-asr | 42.04% | 15.57% | -| facebook/wav2vec2-base-10k-voxpopuli-ft-en | 45.19% | 20.32% | - ---- - -**TIMIT** - -| Model | WER | CER | -| ------------- | ------------- | ------------- | -| facebook/wav2vec2-large-960h-lv60-self | **5.17%** | **1.33%** | -| facebook/wav2vec2-large-960h-lv60 | 6.24% | 1.54% | -| jonatasgrosman/wav2vec2-large-xlsr-53-english | 6.81% | 2.02% | -| facebook/wav2vec2-large-960h | 9.63% | 2.19% | -| facebook/wav2vec2-base-960h | 11.48% | 2.76% | -| elgeish/wav2vec2-large-lv60-timit-asr | 13.83% | 4.36% | -| jonatasgrosman/wav2vec2-large-english | 13.91% | 4.01% | -| facebook/wav2vec2-base-100h | 16.75% | 4.79% | -| elgeish/wav2vec2-base-timit-asr | 25.40% | 8.16% | -| boris/xlsr-en-punctuation | 25.93% | 9.99% | -| facebook/wav2vec2-base-10k-voxpopuli-ft-en | 51.08% | 19.84% | diff --git a/config.json b/config.json index b580e68..8f6c692 100644 --- a/config.json +++ b/config.json @@ -71,6 +71,6 @@ "num_feat_extract_layers": 7, "num_hidden_layers": 24, "pad_token_id": 0, - "transformers_version": "4.5.0.dev0", + "transformers_version": "4.7.0.dev0", "vocab_size": 33 } diff --git a/preprocessor_config.json b/preprocessor_config.json index 0886a48..73caa15 100644 --- a/preprocessor_config.json +++ b/preprocessor_config.json @@ -1,5 +1,6 @@ { "do_normalize": true, + "feature_extractor_type": "Wav2Vec2FeatureExtractor", "feature_size": 1, "padding_side": "right", "padding_value": 0.0, diff --git a/pytorch_model.bin b/pytorch_model.bin index d8c82e8..d2aabbf 100644 --- a/pytorch_model.bin +++ b/pytorch_model.bin @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:f9a3cc8402adf407944f67025114dd0f02af9584bfb266c031f586d601719b0a +oid sha256:7b7688644eeefe1f5760bb4c4a61d085793a3740159fdbf19fd37c5d4f3729bf size 1262069143