Use Silero VAD in Batched Mode (#936)

Replace Pyannote VAD with Silero to reduce code duplication and requirements

MANIFEST.in

@@ -1,4 +1,4 @@
-include faster_whisper/assets/silero_vad.onnx
+include faster_whisper/assets/silero_encoder_v5.onnx
+include faster_whisper/assets/silero_decoder_v5.onnx
 include requirements.txt
 include requirements.conversion.txt
-include faster_whisper/assets/pyannote_vad_model.bin

README.md

@@ -178,9 +178,6 @@ language_info = model.detect_language_multi_segment("audio.mp3")
 
 ### Batched faster-whisper
 
-
-The batched version of faster-whisper is inspired by [whisper-x](https://github.com/m-bain/whisperX) licensed under the BSD-2 Clause license and integrates its VAD model to this library. We modify this implementation and also replaced the feature extraction with a faster torch-based implementation. Batched version improves the speed upto 10-12x compared to openAI implementation and 3-4x compared to the sequential faster_whisper version. It works by transcribing semantically meaningful audio chunks as batches leading to faster inference. 
-
 The following code snippet illustrates how to run inference with batched version on an example audio file. Please also refer to the test scripts of batched faster whisper.
 
 ```python

benchmark/evaluate_yt_commons.py

@@ -0,0 +1,83 @@
+import argparse
+import json
+import os
+
+from io import BytesIO
+
+from datasets import load_dataset
+from evaluate import load
+from pytubefix import YouTube
+from torch.utils.data import DataLoader
+from tqdm import tqdm
+from transformers.models.whisper.english_normalizer import EnglishTextNormalizer
+
+from faster_whisper import BatchedInferencePipeline, WhisperModel, decode_audio
+
+
+def url_to_audio(row):
+    buffer = BytesIO()
+    yt = YouTube(row["link"])
+    video = (
+        yt.streams.filter(only_audio=True, mime_type="audio/mp4")
+        .order_by("bitrate")
+        .desc()
+        .first()
+    )
+    video.stream_to_buffer(buffer)
+    buffer.seek(0)
+    row["audio"] = decode_audio(buffer)
+    return row
+
+
+parser = argparse.ArgumentParser(description="WER benchmark")
+parser.add_argument(
+    "--audio_numb",
+    type=int,
+    default=None,
+    help="Specify the number of validation audio files in the dataset."
+    " Set to None to retrieve all audio files.",
+)
+args = parser.parse_args()
+
+# define the evaluation metric
+wer_metric = load("wer")
+
+with open(os.path.join(os.path.dirname(__file__), "normalizer.json"), "r") as f:
+    normalizer = EnglishTextNormalizer(json.load(f))
+
+dataset = load_dataset("mobiuslabsgmbh/youtube-commons-asr-eval", streaming=True).map(
+    url_to_audio
+)
+dataset = iter(
+    DataLoader(dataset["test"], batch_size=1, prefetch_factor=4, num_workers=2)
+)
+
+model = WhisperModel("large-v3", device="cuda")
+pipeline = BatchedInferencePipeline(model, device="cuda")
+
+
+all_transcriptions = []
+all_references = []
+# iterate over the dataset and run inference
+for i, row in tqdm(enumerate(dataset), desc="Evaluating..."):
+    result, info = pipeline.transcribe(
+        row["audio"][0],
+        batch_size=8,
+        word_timestamps=False,
+        without_timestamps=True,
+    )
+
+    all_transcriptions.append("".join(segment.text for segment in result))
+    all_references.append(row["text"][0])
+    if args.audio_numb and i == (args.audio_numb - 1):
+        break
+
+# normalize predictions and references
+all_transcriptions = [normalizer(transcription) for transcription in all_transcriptions]
+all_references = [normalizer(reference) for reference in all_references]
+
+# compute the WER metric
+wer = 100 * wer_metric.compute(
+    predictions=all_transcriptions, references=all_references
+)
+print("WER: %.3f" % wer)

benchmark/requirements.benchmark.txt

@@ -4,3 +4,4 @@ evaluate
 datasets
 memory_profiler
 py3nvml
+pytubefix