Feat/parse audio (#10133)

### What problem does this PR solve?

Dataflow support audio.  And fix giteeAI's sequence2text model. 

### Type of change

- [x] Bug Fix (non-breaking change which fixes an issue)
- [x] New Feature (non-breaking change which adds functionality)

deepdoc/parser/pdf_parser.py

@@ -74,10 +74,10 @@ class RAGFlowPdfParser:
             recognizer_domain = "layout"
 
         if layout_recognizer_type == "ascend":
-            logging.debug("Using Ascend LayoutRecognizer", flush=True)
+            logging.debug("Using Ascend LayoutRecognizer")
             self.layouter = AscendLayoutRecognizer(recognizer_domain)
         else:  # onnx
-            logging.debug("Using Onnx LayoutRecognizer", flush=True)
+            logging.debug("Using Onnx LayoutRecognizer")
             self.layouter = LayoutRecognizer(recognizer_domain)
         self.tbl_det = TableStructureRecognizer()
 

rag/flow/parser/parser.py

@@ -54,7 +54,9 @@ class ParserParam(ProcessParamBase):
                 "text",
                 "json"
             ],
-            "audio": [],
+            "audio": [
+                "json"
+            ],
             "video": [],
         }
 
@@ -102,7 +104,26 @@ class ParserParam(ProcessParamBase):
                 ],
                 "output_format": "json",
             },
-            "audio": {},
+            "audio": {
+                "suffix":[
+                    "da",
+                    "wave",
+                    "wav",
+                    "mp3",
+                    "aac",
+                    "flac",
+                    "ogg",
+                    "aiff",
+                    "au",
+                    "midi",
+                    "wma",
+                    "realaudio",
+                    "vqf",
+                    "oggvorbis",
+                    "ape"
+                ],
+                "output_format": "json",
+            },
             "video": {},
         }
 
@@ -146,6 +167,12 @@ class ParserParam(ProcessParamBase):
             text_output_format = text_config.get("output_format", "")
             self.check_valid_value(text_output_format, "Text output format abnormal.", self.allowed_output_format["text"])
 
+        audio_config = self.setups.get("audio", "")
+        if audio_config:
+            self.check_empty(audio_config.get("llm_id"), "VLM")
+            audio_language = audio_config.get("lang", "")
+            self.check_empty(audio_language, "Language")
+
     def get_input_form(self) -> dict[str, dict]:
         return {}
 
@@ -313,6 +340,30 @@ class Parser(ProcessBase):
 
         self.set_output("text", txt)
 
+    def _audio(self, from_upstream: ParserFromUpstream):
+        import os
+        import tempfile
+
+        self.callback(random.randint(1, 5) / 100.0, "Start to work on an audio.")
+
+        blob = from_upstream.blob
+        name = from_upstream.name
+        conf = self._param.setups["audio"]
+        self.set_output("output_format", conf["output_format"])
+
+        lang = conf["lang"]
+        _, ext = os.path.splitext(name)
+        tmp_path = ""
+        with tempfile.NamedTemporaryFile(suffix=ext) as tmpf:
+            tmpf.write(blob)
+            tmpf.flush()
+            tmp_path = os.path.abspath(tmpf.name)
+
+            seq2txt_mdl = LLMBundle(self._canvas.get_tenant_id(), LLMType.SPEECH2TEXT, lang=lang)
+            txt = seq2txt_mdl.transcription(tmp_path)
+
+            self.set_output("text", txt)
+
     async def _invoke(self, **kwargs):
         function_map = {
             "pdf": self._pdf,
@@ -321,6 +372,7 @@ class Parser(ProcessBase):
             "word": self._word,
             "text": self._text,
             "image": self._image,
+            "audio": self._audio,
         }
         try:
             from_upstream = ParserFromUpstream.model_validate(kwargs)

rag/flow/tests/dsl_examples/general_pdf_all.json

@@ -60,6 +60,28 @@
                     "gif"
                   ],
                   "output_format": "text"
+                },
+                "audio": {
+                  "suffix": [
+                    "da",
+                    "wave",
+                    "wav",
+                    "mp3",
+                    "aac",
+                    "flac",
+                    "ogg",
+                    "aiff",
+                    "au",
+                    "midi",
+                    "wma",
+                    "realaudio",
+                    "vqf",
+                    "oggvorbis",
+                    "ape"
+                  ],
+                  "lang": "Chinese",
+                  "llm_id": "SenseVoiceSmall",
+                  "output_format": "json"
                 }
               }
           }

rag/llm/sequence2txt_model.py

@@ -218,7 +218,7 @@ class GPUStackSeq2txt(Base):
 class GiteeSeq2txt(Base):
     _FACTORY_NAME = "GiteeAI"
 
-    def __init__(self, key, model_name="whisper-1", base_url="https://ai.gitee.com/v1/"):
+    def __init__(self, key, model_name="whisper-1", base_url="https://ai.gitee.com/v1/", **kwargs):
         if not base_url:
             base_url = "https://ai.gitee.com/v1/"
         self.client = OpenAI(api_key=key, base_url=base_url)