[feat](audio): create endpoint

fix(audio): config && create audio endpoint

fix l'API audio et ajout d'un fichier pour tester

clean clode

fix(audio): make it work

app/endpoints/audio.py

@@ -0,0 +1,55 @@
+from typing import Optional, Union
+
+from fastapi import APIRouter, Form, Security, HTTPException, UploadFile, File
+
+from app.schemas.audio import AudioTranscription, AudioTranscriptionVerbose
+from app.schemas.config import AUDIO_MODEL_TYPE
+from app.utils.security import check_api_key
+from app.utils.lifespan import clients
+
+
+router = APIRouter()
+
+
+@router.post("/audio/transcriptions")
+async def audio_transcriptions(
+    file: UploadFile = File(...),
+    model: str = Form(...),
+    prompt: Optional[str] = Form(None),
+    response_format: str = Form(...),
+    temperature: float = Form(0.2),
+    _: str = Security(check_api_key),
+) -> Union[AudioTranscription, AudioTranscriptionVerbose]:
+    """
+    API de transcription similaire à l'API d'OpenAI.
+    """
+    client = clients.models[model]
+
+    if client.type != AUDIO_MODEL_TYPE:
+        raise HTTPException(status_code=400, detail="Le modèle n'est pas un modèle audio.")
+
+    url = f"{client.base_url}audio/transcriptions"
+    headers = {"Authorization": f"Bearer {client.api_key}"}
+
+    """
+    with httpx.AsyncClient(timeout=20) as async_client:
+        response = await async_client.request(
+            method="POST",
+            url=url,
+            headers=headers,
+            files={"file": file.file},
+            json={"prompt": prompt, "response_format": response_format, "temperature": temperature},
+        )
+        response.raise_for_status()
+   
+        data = response.json()
+        return data
+        if response_format == "verbon_json":
+            return AudioTranscriptionVerbose(**data)
+        if response_format == "json":
+            return AudioTranscription(**data)
+        raise HTTPException(status_code=400, detail="Invalid response format")
+    """
+    file_content = await file.read()
+    response = client.audio.transcriptions.create(file=("audio.mp3", file_content, file.content_type), model=model)
+    return response

app/helpers/_modelclients.py

@@ -10,7 +10,7 @@
 from app.schemas.models import Model, Models
 from app.utils.config import LOGGER
 from app.utils.exceptions import ContextLengthExceededException, ModelNotAvailableException, ModelNotFoundException
-from app.utils.variables import EMBEDDINGS_MODEL_TYPE, LANGUAGE_MODEL_TYPE
+from app.utils.variables import EMBEDDINGS_MODEL_TYPE, LANGUAGE_MODEL_TYPE, AUDIO_MODEL_TYPE
 
 
 def get_models_list(self, *args, **kwargs):
@@ -45,6 +45,15 @@ def get_models_list(self, *args, **kwargs):
             self.created = round(time.time())
             self.max_model_len = response.get("max_input_length", None)
 
+        elif self.type == AUDIO_MODEL_TYPE:
+            endpoint = f"{self.base_url}models/Systran/faster-whisper-large-v3"
+            response = requests.get(url=endpoint, headers=headers, timeout=self.DEFAULT_TIMEOUT).json()
+
+            self.id = response["id"]
+            self.owned_by = response.get("owned_by", "")
+            self.created = response.get("created", round(time.time()))
+            self.max_model_len = response.get("max_model_len", None)
+
         self.status = "available"
 
     except Exception:
@@ -61,6 +70,9 @@ def get_models_list(self, *args, **kwargs):
     )
 
     return Models(data=[data])
+
+
+    return Models(data=data)
 
 
 def check_context_length(self, messages: List[Dict[str, str]], add_special_tokens: bool = True):
@@ -99,11 +111,12 @@ def create_embeddings(self, *args, **kwargs):
 class ModelClient(OpenAI):
     DEFAULT_TIMEOUT = 10
 
-    def __init__(self, type=Literal[EMBEDDINGS_MODEL_TYPE, LANGUAGE_MODEL_TYPE], *args, **kwargs):
+    def __init__(self, type=Literal[EMBEDDINGS_MODEL_TYPE, LANGUAGE_MODEL_TYPE, AUDIO_MODEL_TYPE], *args, **kwargs):
         """
         ModelClient class extends OpenAI class to support custom methods.
         """
-        super().__init__(timeout=self.DEFAULT_TIMEOUT, *args, **kwargs)
+        timeout = 60 if type == AUDIO_MODEL_TYPE else self.DEFAULT_TIMEOUT
+        super().__init__(timeout=timeout, *args, **kwargs)
         self.type = type
 
         # set attributes for unavailable models

app/main.py

@@ -1,12 +1,13 @@
 from fastapi import FastAPI, Response, Security
 
-from app.endpoints import chat, chunks, collections, completions, embeddings, files, models, search, documents
+from app.endpoints import audio, chat, chunks, collections, completions, embeddings, files, models, search, documents
 from app.helpers import ContentSizeLimitMiddleware
 from app.schemas.security import User
 from app.utils.config import APP_CONTACT_EMAIL, APP_CONTACT_URL, APP_DESCRIPTION, APP_VERSION
 from app.utils.lifespan import lifespan
 from app.utils.security import check_api_key
 
+
 app = FastAPI(
     title="Albert API",
     version=APP_VERSION,
@@ -32,6 +33,7 @@ def health(user: User = Security(check_api_key)):
 app.include_router(chat.router, tags=["Core"], prefix="/v1")
 app.include_router(completions.router, tags=["Core"], prefix="/v1")
 app.include_router(embeddings.router, tags=["Core"], prefix="/v1")
+app.include_router(audio.router, tags=["Core"], prefix="/v1")
 
 # RAG
 app.include_router(search.router, tags=["Retrieval Augmented Generation"], prefix="/v1")

app/schemas/audio.py

@@ -0,0 +1,40 @@
+from typing import Dict, List, Optional, Union, Literal, Tuple
+
+from pydantic import BaseModel
+
+
+#cf https://platform.openai.com/docs/api-reference/audio/createTranscription
+class AudioTranscriptionRequest(BaseModel):
+    model: Optional[str] = "Systran/faster-distil-whisper-large-v3"
+    language: Optional[str] = None
+    prompt: Optional[str] = None
+    response_format: Optional[str] = "json"
+    temperature: Optional[float] = 0
+    timestamp: Optional[bool] = False
+
+class AudioTranscription(BaseModel):
+    text: str
+
+class Word(BaseModel):
+    word: str
+    start: float
+    end: float
+
+class Segment(BaseModel):
+    id: int
+    seek: int
+    start: float
+    end: float
+    text: str
+    tokens: List[int]
+    temperature: float
+    avg_logprob: float
+    compression_ratio: float
+    no_speech_prob: float
+
+class AudioTranscriptionVerbose(BaseModel):
+    language: str
+    duration: float
+    text: str
+    words: List[Word]
+    segments: List[Segment]

app/schemas/config.py

@@ -2,7 +2,7 @@
 
 from pydantic import BaseModel, Field, model_validator
 
-from app.utils.variables import EMBEDDINGS_MODEL_TYPE, LANGUAGE_MODEL_TYPE
+from app.utils.variables import EMBEDDINGS_MODEL_TYPE, LANGUAGE_MODEL_TYPE, AUDIO_MODEL_TYPE
 
 
 class Key(BaseModel):
@@ -16,7 +16,7 @@ class Auth(BaseModel):
 
 class Model(BaseModel):
     url: str
-    type: Literal[LANGUAGE_MODEL_TYPE, EMBEDDINGS_MODEL_TYPE]
+    type: Literal[LANGUAGE_MODEL_TYPE, EMBEDDINGS_MODEL_TYPE, AUDIO_MODEL_TYPE]
     key: Optional[str] = "EMPTY"
     search_internet: bool = False
 

app/schemas/models.py

@@ -3,11 +3,11 @@
 from openai.types import Model
 from pydantic import BaseModel
 
-from app.utils.variables import EMBEDDINGS_MODEL_TYPE, LANGUAGE_MODEL_TYPE
+from app.utils.variables import EMBEDDINGS_MODEL_TYPE, LANGUAGE_MODEL_TYPE, AUDIO_MODEL_TYPE
 
 
 class Model(Model):
-    type: Literal[LANGUAGE_MODEL_TYPE, EMBEDDINGS_MODEL_TYPE]
+    type: Literal[LANGUAGE_MODEL_TYPE, EMBEDDINGS_MODEL_TYPE, AUDIO_MODEL_TYPE]
     status: Literal["available", "unavailable"]
 
 

app/tests/test_audio.py

@@ -0,0 +1,21 @@
+from openai import OpenAI
+import requests
+
+with open('audio.mp3', 'wb') as fopen:
+    r = requests.get('https://huggingface.co/datasets/huseinzol05/temp-storage/resolve/main/Lex-Fridman-on-Grigori-Perelman-turning-away-1million-and-Fields-Medal.mp3?download=true')
+    fopen.write(r.content)
+
+client = OpenAI(
+    api_key='jules-f0b946fa-0da4-4500-a536-e4fb7448d6bc',
+    base_url = 'http://127.0.0.1:8080'
+)
+
+audio_file = open("audio.mp3", "rb")
+transcript = client.audio.transcriptions.create(
+  file=audio_file,
+  model="Systran/faster-distil-whisper-large-v3",
+  response_format="json",
+  #timestamp_granularities="segment"
+)
+
+print(transcript)

app/utils/variables.py

@@ -1,3 +1,4 @@
+AUDIO_MODEL_TYPE = "automatic-speech-recognition"
 INTERNET_COLLECTION_ID = "internet"
 PUBLIC_COLLECTION_TYPE = "public"
 PRIVATE_COLLECTION_TYPE = "private"