new structure for transcription engine

- found problems with mps
- integrated jinja2 templates for rendering (.tex?)
- raw ui & bad structure

transcription/preprocessing/audio_preprocessor.py

@@ -0,0 +1,35 @@
+from transcription.audio import Audio
+import torchaudio
+
+class AudioPreprocessor:
+    TARGET_SAMPLING_RATE: int = 16000
+
+    # for different models in future
+    # def __init__(self, model):
+    #     pass
+
+    def _resample(
+        self,
+        audio: Audio
+    ) -> None:
+        if audio.sr != self.TARGET_SAMPLING_RATE:
+            audio.waveform = torchaudio.functional.resample(
+                audio.waveform, 
+                audio.sr, 
+                self.TARGET_SAMPLING_RATE
+            )
+    
+    def _to_mono(
+        self,
+        audio: Audio
+    ) -> None:
+        if audio.waveform.shape[0] > 1:
+            audio.waveform = audio.waveform.mean(dim=0, keepdim=True)
+        audio.waveform = audio.waveform.squeeze(0)
+    
+    def prepare(
+        self, 
+        audio: Audio
+    ):
+        self._resample(audio)
+        self._to_mono(audio)

transcription/preprocessing/splitter.py

@@ -0,0 +1,50 @@
+import torch
+from typing import List
+
+class Splitter:
+    def __init__(
+        self,
+        chunkSize: int,
+        batchSize: int,
+    ) -> None:
+        self.chunkSize = chunkSize * 16000 # 16 kHz after resampling
+        self.batchSize = batchSize
+
+    # maybe raise some exceptions here?
+    def _split_to_chunks(
+        self, 
+        waveform: torch.Tensor,
+    ) -> List:
+        totalSamples = waveform.shape[0]
+        chunksCount = (totalSamples + self.chunkSize - 1) // self.chunkSize
+
+        chunks: List = []
+        # tqdm or something here? 
+        for chunkNum in range(chunksCount):
+            start = chunkNum * self.chunkSize
+            end = min((chunkNum + 1) * self.chunkSize, totalSamples)
+
+            chunk = waveform[start : end].cpu().numpy().astype("float32")
+            chunks.append(chunk)
+        
+        return chunks
+
+    def _split_to_batches(
+        self,
+        chunks: List,
+    ) -> List:
+        batches: List = []
+
+        for i in range(0, len(chunks), self.batchSize):
+            batch = chunks[i : i + self.batchSize]
+            batches.append(batch)
+        
+        return batches
+
+    def split(
+        self,
+        waveform: torch.Tensor 
+    ) -> List:
+        chunks = self._split_to_chunks(waveform)
+        batches = self._split_to_batches(chunks)
+        return batches