Add the initial_prompt parameter (#2)

* Add the initial_prompt parameter

* Add docstring

faster_whisper/transcribe.py

@@ -32,6 +32,7 @@ class TranscriptionOptions(
             "compression_ratio_threshold",
             "condition_on_previous_text",
             "temperatures",
+            "initial_prompt",
         ),
     )
 ):
@@ -64,16 +65,9 @@ class WhisperModel:
         )
 
         self.feature_extractor = FeatureExtractor()
-        self.decoder = tokenizers.decoders.ByteLevel()
+        self.tokenizer = tokenizers.Tokenizer.from_pretrained("openai/whisper-tiny")
-
+        self.eot_id = self.tokenizer.token_to_id("<|endoftext|>")
-        with open(os.path.join(model_path, "vocabulary.txt")) as vocab_file:
+        self.timestamp_begin_id = self.tokenizer.token_to_id("<|notimestamps|>") + 1
-            self.ids_to_tokens = [line.rstrip("\n") for line in vocab_file]
-            self.tokens_to_ids = {
-                token: i for i, token in enumerate(self.ids_to_tokens)
-            }
-
-        self.eot_id = self.tokens_to_ids["<|endoftext|>"]
-        self.timestamp_begin_id = self.tokens_to_ids["<|notimestamps|>"] + 1
         self.input_stride = 2
         self.time_precision = 0.02
         self.max_length = 448
@@ -90,6 +84,7 @@ class WhisperModel:
         log_prob_threshold=-1.0,
         no_speech_threshold=0.6,
         condition_on_previous_text=True,
+        initial_prompt=None,
     ):
         """Transcribes an input file.
 
@@ -114,6 +109,7 @@ class WhisperModel:
             as a prompt for the next window; disabling may make the text inconsistent across
             windows, but the model becomes less prone to getting stuck in a failure loop,
             such as repetition looping or timestamps going out of sync.
+          initial_prompt: Optional text to provide as a prompt for the first window.
 
         Returns:
           A tuple with:
@@ -146,6 +142,7 @@ class WhisperModel:
             temperatures=(
                 temperature if isinstance(temperature, (list, tuple)) else [temperature]
             ),
+            initial_prompt=initial_prompt,
         )
 
         segments = self.generate_segments(features, language, options)
@@ -179,6 +176,13 @@ class WhisperModel:
         all_tokens = []
         prompt_reset_since = 0
 
+        if options.initial_prompt is not None:
+            initial_prompt = " " + options.initial_prompt.strip()
+            initial_prompt_tokens = self.tokenizer.encode(
+                initial_prompt, add_special_tokens=False
+            )
+            all_tokens.extend(initial_prompt_tokens.ids)
+
         while offset < num_frames:
             time_offset = offset * self.feature_extractor.time_per_frame
             segment = self.get_segment(features, offset)
@@ -253,11 +257,8 @@ class WhisperModel:
                 prompt_reset_since = len(all_tokens)
 
     def decode_text_tokens(self, tokens):
-        text_tokens = [
+        text_tokens = [token for token in tokens if token < self.eot_id]
-            self.ids_to_tokens[token] for token in tokens if token < self.eot_id
+        return self.tokenizer.decode(text_tokens)
-        ]
-
-        return self.decoder.decode(text_tokens)
 
     def generate_with_fallback(self, segment, prompt, options):
         features = self.get_input(segment)
@@ -304,13 +305,13 @@ class WhisperModel:
         prompt = []
 
         if previous_tokens:
-            prompt.append(self.tokens_to_ids["<|startofprev|>"])
+            prompt.append(self.tokenizer.token_to_id("<|startofprev|>"))
             prompt.extend(previous_tokens[-(self.max_length // 2 - 1) :])
 
         prompt += [
-            self.tokens_to_ids["<|startoftranscript|>"],
+            self.tokenizer.token_to_id("<|startoftranscript|>"),
-            self.tokens_to_ids["<|%s|>" % language],
+            self.tokenizer.token_to_id("<|%s|>" % language),
-            self.tokens_to_ids["<|transcribe|>"],
+            self.tokenizer.token_to_id("<|transcribe|>"),
         ]
 
         return prompt