Allow Whisper decoding to use the full decoder context

python/tests/test_transformers.py

@@ -864,6 +864,58 @@ def _get_features(audio):
             transcription = processor.decode(token_ids)
             assert transcription == expected_transcription
 
+    @test_utils.only_on_linux
+    def test_transformers_whisper_full_context(self, tmp_dir):
+        import transformers
+
+        model_name = "openai/whisper-tiny"
+        converter = ctranslate2.converters.TransformersConverter(model_name)
+        output_dir = str(tmp_dir.join("ctranslate2_model"))
+        output_dir = converter.convert(output_dir)
+
+        audio_path = os.path.join(test_utils.get_data_dir(), "audio", "jfk.npy")
+        audio = np.load(audio_path)
+
+        processor = transformers.WhisperProcessor.from_pretrained(model_name)
+        # Pad after computing the log-Mel spectrogram to match the openai/whisper behavior.
+        inputs = processor(audio, padding=False, sampling_rate=16000)
+        features = inputs.input_features[0]
+        features = np.pad(features, [(0, 0), (0, 3000 - features.shape[-1])])
+        features = ctranslate2.StorageView.from_array(np.expand_dims(features, 0))
+
+        model = ctranslate2.models.Whisper(output_dir)
+
+        # Suppress <|endoftext|> so generation always runs until max_length is
+        # reached, regardless of the actual audio content.
+        eot_id = processor.tokenizer.convert_tokens_to_ids("<|endoftext|>")
+
+        def _generated_length(prompt):
+            result = model.generate(
+                features,
+                [prompt],
+                beam_size=1,
+                num_hypotheses=1,
+                suppress_tokens=[eot_id],
+            )[0]
+            return len(result.sequences_ids[0])
+
+        no_timestamps_prompt = [
+            "<|startoftranscript|>",
+            "<|en|>",
+            "<|transcribe|>",
+            "<|notimestamps|>",
+        ]
+        timestamped_prompt = ["<|startoftranscript|>", "<|en|>", "<|transcribe|>"]
+
+        # With the default max_length of 448, decoding can use the entire
+        # remaining decoder context (448 minus the prompt length), no longer
+        # capped to half of max_length. This applies to both no-timestamp and
+        # timestamped decoding, since the previous 224-token cap was not
+        # actually related to timestamps. The timestamped prompt is one
+        # token shorter, so it gets one extra position of context.
+        assert _generated_length(no_timestamps_prompt) == 445
+        assert _generated_length(timestamped_prompt) == 446
+
     @test_utils.only_on_linux
     @test_utils.on_available_devices
     @pytest.mark.parametrize(

src/models/whisper.cc

@@ -300,7 +300,7 @@ namespace ctranslate2 {
       decoding_options.length_penalty = options.length_penalty;
       decoding_options.repetition_penalty = options.repetition_penalty;
       decoding_options.no_repeat_ngram_size = options.no_repeat_ngram_size;
-      decoding_options.max_length = std::min(total_max_length / 2, total_max_length - start_step);
+      decoding_options.max_length = total_max_length - start_step;
       decoding_options.sampling_topk = options.sampling_topk;
       decoding_options.sampling_temperature = options.sampling_temperature;
       decoding_options.num_hypotheses = options.num_hypotheses;