From f17baa7d71a6dda04df67b1676fcabed9f35d2c6 Mon Sep 17 00:00:00 2001
From: sssshhhhhh <193317444+sssshhhhhh@users.noreply.github.com>
Date: Sat, 31 Jan 2026 22:48:01 +1100
Subject: [PATCH] Assume less about whisper vocab

---
 src/models/whisper.cc | 7 +++++--
 1 file changed, 5 insertions(+), 2 deletions(-)

diff --git a/src/models/whisper.cc b/src/models/whisper.cc
index 7b071ace6..8c3612f1e 100644
--- a/src/models/whisper.cc
+++ b/src/models/whisper.cc
@@ -69,9 +69,12 @@ namespace ctranslate2 {
       _no_speech_id = vocabulary.to_id("<|nospeech|>");
       if (_no_speech_id == vocabulary.unk_id())
         _no_speech_id = vocabulary.to_id("<|nocaptions|>");
-      _is_multilingual = vocabulary.size() >= 51865;
+      _is_multilingual = vocabulary.to_id("") != vocabulary.unk_id();
       _n_mels = _encoder->input_size();
-      _num_languages = vocabulary.size() - 51765 - (_is_multilingual ? 1 : 0);
+      // vocab: text tokens..., <|endoftext|>, <|startoftranscript|>,
+      // lang tokens..., <|translate|>, <|transcribe|>, <|startoflm|>,
+      // <|startofprev|>, <|nospeech|>, <|notimestamps|>, time tokens...
+      _num_languages = _no_speech_id - _sot_id - 5;
     }
 
     StorageView WhisperReplica::encode(StorageView features, const bool to_cpu) {