Add custom vocabulary / hotwords support

whisper_live/backend/faster_whisper_backend.py

@@ -34,6 +34,7 @@ class ServeClientFasterWhisper(ServeClientBase):
         same_output_threshold=7,
         cache_path="~/.cache/whisper-live/",
         translation_queue=None,
+        hotwords=None,
     ):
         """
         Initialize a ServeClient instance.
@@ -78,6 +79,7 @@ class ServeClientFasterWhisper(ServeClientBase):
         self.task = task
         self.initial_prompt = initial_prompt
         self.vad_parameters = vad_parameters or {"threshold": 0.5}
+        self.hotwords = hotwords
 
         device = "cuda" if torch.cuda.is_available() else "cpu"
         if device == "cuda":
@@ -231,7 +233,8 @@ class ServeClientFasterWhisper(ServeClientBase):
             language=self.language,
             task=self.task,
             vad_filter=self.use_vad,
-            vad_parameters=self.vad_parameters if self.use_vad else None)
+            vad_parameters=self.vad_parameters if self.use_vad else None,
+            hotwords=self.hotwords)
         if ServeClientFasterWhisper.SINGLE_MODEL:
             ServeClientFasterWhisper.SINGLE_MODEL_LOCK.release()
 

whisper_live/client.py

@@ -43,6 +43,7 @@ class Client:
         translation_srt_file_path="output_translated.srt",
         enable_timestamps=False,
         display_segments=4,
+        hotwords=None,
     ):
         """
         Initializes a Client instance for audio recording and streaming to a server.
@@ -101,6 +102,7 @@ class Client:
             self.task = "translate"
         self.enable_timestamps = enable_timestamps
         self.display_segments = display_segments
+        self.hotwords = hotwords
 
         self.audio_bytes = None
 
@@ -299,6 +301,7 @@ class Client:
                     "same_output_threshold": self.same_output_threshold,
                     "enable_translation": self.enable_translation,
                     "target_language": self.target_language,
+                    "hotwords": self.hotwords,
                 }
             )
         )
@@ -820,6 +823,7 @@ class TranscriptionClient(TranscriptionTeeClient):
         translation_srt_file_path="./output_translated.srt",
         enable_timestamps=False,
         display_segments=4,
+        hotwords=None,
     ):
         self.client = Client(
             host,
@@ -842,6 +846,7 @@ class TranscriptionClient(TranscriptionTeeClient):
             translation_srt_file_path=translation_srt_file_path,
             enable_timestamps=enable_timestamps,
             display_segments=display_segments,
+            hotwords=hotwords,
         )
 
         if save_output_recording and not output_recording_filename.endswith(".wav"):

whisper_live/server.py

@@ -288,7 +288,8 @@ class TranscriptionServer:
                     clip_audio=options.get("clip_audio", False),
                     same_output_threshold=options.get("same_output_threshold", 10),
                     cache_path=self.cache_path,
-                    translation_queue=translation_queue
+                    translation_queue=translation_queue,
+                    hotwords=options.get("hotwords"),
                 )
 
                 logging.info("Running faster_whisper backend.")
@@ -533,7 +534,8 @@ class TranscriptionServer:
                 include: Optional[List[str]] = Form(default=None),
                 known_speaker_names: Optional[List[str]] = Form(default=None),
                 known_speaker_references: Optional[List[str]] = Form(default=None),
-                stream: bool = Form(default=False)
+                stream: bool = Form(default=False),
+                hotwords: Optional[str] = Form(default=None),
             ):
                 if stream:
                     return JSONResponse({"error": "Streaming not supported in this backend."}, status_code=400)
@@ -564,7 +566,8 @@ class TranscriptionServer:
                         initial_prompt=prompt,
                         temperature=temperature,
                         vad_filter=False,
-                        word_timestamps=(timestamp_granularities and "word" in timestamp_granularities)
+                        word_timestamps=(timestamp_granularities and "word" in timestamp_granularities),
+                        hotwords=hotwords,
                     )
 
                     text = " ".join([s.text.strip() for s in segments])