Add vad_filter to ctranslate transcriber

Helps us to correctly detect the language if there is no voice in the
first 30 seconds

Also helps to lower hallucinations

packages/tests/src/shared/transcription.ts

@@ -28,7 +28,11 @@ export function getCustomModelPath (modelName: CustomModelName) {
 
 // ---------------------------------------------------------------------------
 
-export async function checkAutoCaption (servers: PeerTubeServer[], uuid: string, captionContains = 'WEBVTT\n\n00:00.000 --> 00:') {
+export async function checkAutoCaption (
+  servers: PeerTubeServer[],
+  uuid: string,
+  captionContains = new RegExp('^WEBVTT\\n\\n00:00.\\d{3} --> 00:')
+) {
   for (const server of servers) {
     const body = await server.captions.list({ videoId: uuid })
     expect(body.total).to.equal(1)

packages/transcription/src/whisper/transcriber/ctranslate2-transcriber.ts

@@ -35,6 +35,8 @@ export class Ctranslate2Transcriber extends OpenaiTranscriber {
       ...modelArgs,
       '--word_timestamps',
       'True',
+      '--vad_filter',
+      'true',
       '--output_format',
       'all',
       '--output_dir',