1.remove chunk_length_s=30 2.set task to translate for genai

tools/llm_bench/llm_bench_utils/ov_utils.py

@@ -12,6 +12,7 @@
 from llm_bench_utils.hook_common import get_bench_hook
 from llm_bench_utils.config_class import OV_MODEL_CLASSES_MAPPING, TOKENIZE_CLASSES_MAPPING, DEFAULT_MODEL_CLASSES
 import openvino.runtime.opset13 as opset
+from transformers import pipeline
 
 
 def generate_simplified(self, *args, **kwargs):
@@ -324,12 +325,9 @@ def create_speech_2txt_model(model_path, device, **kwargs):
     from_pretrained_time = end - start
     log.info(f'From pretrained time: {from_pretrained_time:.2f}s')
     processor = AutoProcessor.from_pretrained(model_path)
-
-    from transformers import pipeline
     pipe = pipeline(
         "automatic-speech-recognition",
         model=ov_model,
-        chunk_length_s=30,
         tokenizer=processor.tokenizer,
         feature_extractor=processor.feature_extractor
     )

tools/llm_bench/task/speech_to_text_generation.py

@@ -46,7 +46,7 @@ def run_speech_2_txt_generation(input_param, args, md5_list, iter_data_list):
             max_new_tokens=max_gen_tokens,
             # 'task' and 'language' parameters are supported for multilingual models only
             language=speech_language,
-            task="transcribe",
+            task="translate",
             return_timestamps=ret_timestamps
         )
         end = time.perf_counter()
@@ -55,13 +55,18 @@ def run_speech_2_txt_generation(input_param, args, md5_list, iter_data_list):
         result_text = result_text.texts[0]
     else:
         start = time.perf_counter()
-        result_text = pipe(raw_speech, generate_kwargs={"task": 'translate'}, return_timestamps=ret_timestamps)["text"]
+        result_text = pipe(
+            raw_speech,
+            generate_kwargs={"task": 'translate', "language": speech_language},
+            return_timestamps=ret_timestamps
+        )["text"]
         end = time.perf_counter()
         tm_list = whisper_hook.get_time_list()
         tm_infer_list = whisper_hook.get_time_infer_list()
     log.debug('latency of all tokens:')
     [log.debug('[{}]{:.4f}'.format(idx, tm)) for idx, tm in enumerate(tm_list)]
-
+    log.debug('latency of all infers:')
+    [log.debug('[{}]{:.4f}'.format(idx, tm)) for idx, tm in enumerate(tm_infer_list)]
     generation_time = end - start
     out_data = processor.tokenizer(result_text, return_tensors='pt')
     out_tokens = out_data['input_ids'] if 'input_ids' in out_data else out_data