Add support for faster-whisper multilingual option

The faster-whisper multilingual option allows language detection to be performed
on each segment. With the multilingual option enabled, you can get better
transcription results if the language was incorrectly detected in the first
place, or if the spoken language is switched in the audio.

src/whisper_ctranslate2/commandline.py

@@ -364,6 +364,13 @@ def read_command_line():
             help="When using Batched transcription the maximum number of parallel requests to model for decoding.",
         )
 
+        algorithm_args.add_argument(
+            "--multilingual",
+            type=CommandLine._str2bool,
+            default=False,
+            help="Perform language detection on every segment",
+        )
+
         vad_args = parser.add_argument_group("VAD filter arguments")
 
         vad_args.add_argument(

src/whisper_ctranslate2/transcribe.py

@@ -52,6 +52,7 @@ class TranscriptionOptions(NamedTuple):
     vad_min_speech_duration_ms: Optional[int]
     vad_max_speech_duration_s: Optional[int]
     vad_min_silence_duration_ms: Optional[int]
+    multilingual: bool
 
 
 class Transcribe:
@@ -179,6 +180,7 @@ def inference(
             vad_filter=vad,
             vad_parameters=vad_parameters,
             **batch_size,
+            multilingual=options.multilingual,
         )
 
         language_name = LANGUAGES[info.language].title()

src/whisper_ctranslate2/whisper_ctranslate2.py

@@ -74,6 +74,7 @@ def get_transcription_options(args):
         vad_min_speech_duration_ms=args.pop("vad_min_speech_duration_ms"),
         vad_max_speech_duration_s=args.pop("vad_max_speech_duration_s"),
         vad_min_silence_duration_ms=args.pop("vad_min_silence_duration_ms"),
+        multilingual=args.pop("multilingual"),
     )