Prepend prefix tokens with the initial timestamp token (#358)

2023-07-18 15:22:39 +02:00
parent 2a37390fed
commit 0e051a5b77
3 changed files with 21 additions and 1 deletions
--- a/faster_whisper/transcribe.py
+++ b/faster_whisper/transcribe.py
@@ -686,6 +686,8 @@ class WhisperModel:
            prefix_tokens = tokenizer.encode(" " + prefix.strip())
            if len(prefix_tokens) >= self.max_length // 2:
                prefix_tokens = prefix_tokens[: self.max_length // 2 - 1]
+            if not without_timestamps:
+                prompt.append(tokenizer.timestamp_begin)
            prompt.extend(prefix_tokens)

        return prompt