openai compatible for asr/tts (#929)

* openai compatible for asr/tts

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* add dep

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

Author: Spycsh

comps/asr/whisper/README.md

@@ -38,6 +38,14 @@ pip install optimum[habana]
 cd dependency/
 nohup python whisper_server.py --device=hpu &
 python check_whisper_server.py
+
+# Or use openai protocol compatible curl command
+# Please refer to https://platform.openai.com/docs/api-reference/audio/createTranscription
+wget https://github.com/intel/intel-extension-for-transformers/raw/main/intel_extension_for_transformers/neural_chat/assets/audio/sample.wav
+curl http://localhost:7066/v1/audio/transcriptions \
+  -H "Content-Type: multipart/form-data" \
+  -F file="@./sample.wav" \
+  -F model="openai/whisper-small"
 ```
 
 ### 1.3 Start ASR Service/Test
@@ -114,6 +122,7 @@ docker run -d -p 9099:9099 --ipc=host -e http_proxy=$http_proxy -e https_proxy=$
 # curl
 http_proxy="" curl http://localhost:9099/v1/audio/transcriptions -XPOST -d '{"byte_str": "UklGRigAAABXQVZFZm10IBIAAAABAAEARKwAAIhYAQACABAAAABkYXRhAgAAAAEA"}' -H 'Content-Type: application/json'
 
+
 # python
 python check_asr_server.py
 ```

comps/asr/whisper/dependency/whisper_model.py

@@ -30,10 +30,10 @@ def __init__(
         from transformers import WhisperForConditionalGeneration, WhisperProcessor
 
         self.device = device
-        asr_model_name_or_path = os.environ.get("ASR_MODEL_PATH", model_name_or_path)
-        print("Downloading model: {}".format(asr_model_name_or_path))
-        self.model = WhisperForConditionalGeneration.from_pretrained(asr_model_name_or_path).to(self.device)
-        self.processor = WhisperProcessor.from_pretrained(asr_model_name_or_path)
+        self.asr_model_name_or_path = os.environ.get("ASR_MODEL_PATH", model_name_or_path)
+        print("Downloading model: {}".format(self.asr_model_name_or_path))
+        self.model = WhisperForConditionalGeneration.from_pretrained(self.asr_model_name_or_path).to(self.device)
+        self.processor = WhisperProcessor.from_pretrained(self.asr_model_name_or_path)
         self.model.eval()
 
         self.language = language

comps/asr/whisper/dependency/whisper_server.py

@@ -5,14 +5,21 @@
 import base64
 import os
 import uuid
+from typing import List, Optional, Union
 
 import uvicorn
-from fastapi import FastAPI, Request
+from fastapi import FastAPI, File, Form, Request, UploadFile
 from fastapi.responses import Response
 from pydub import AudioSegment
 from starlette.middleware.cors import CORSMiddleware
 from whisper_model import WhisperModel
 
+from comps import CustomLogger
+from comps.cores.proto.api_protocol import AudioTranscriptionResponse
+
+logger = CustomLogger("whisper")
+logflag = os.getenv("LOGFLAG", False)
+
 app = FastAPI()
 asr = None
 
@@ -29,7 +36,7 @@ async def health() -> Response:
 
 @app.post("/v1/asr")
 async def audio_to_text(request: Request):
-    print("Whisper generation begin.")
+    logger.info("Whisper generation begin.")
     uid = str(uuid.uuid4())
     file_name = uid + ".wav"
     request_dict = await request.json()
@@ -44,13 +51,58 @@ async def audio_to_text(request: Request):
     try:
         asr_result = asr.audio2text(file_name)
     except Exception as e:
-        print(e)
+        logger.error(e)
         asr_result = e
     finally:
         os.remove(file_name)
     return {"asr_result": asr_result}
 
 
+@app.post("/v1/audio/transcriptions")
+async def audio_transcriptions(
+    file: UploadFile = File(...),  # Handling the uploaded file directly
+    model: str = Form("openai/whisper-small"),
+    language: str = Form("english"),
+    prompt: str = Form(None),
+    response_format: str = Form("json"),
+    temperature: float = Form(0),
+    timestamp_granularities: List[str] = Form(None),
+):
+    logger.info("Whisper generation begin.")
+    audio_content = await file.read()
+    # validate the request parameters
+    if model != asr.asr_model_name_or_path:
+        raise Exception(
+            f"ASR model mismatch! Please make sure you pass --model_name_or_path or set environment variable ASR_MODEL_PATH to {model}"
+        )
+    asr.language = language
+    if prompt is not None or response_format != "json" or temperature != 0 or timestamp_granularities is not None:
+        logger.warning(
+            "Currently parameters 'language', 'response_format', 'temperature', 'timestamp_granularities' are not supported!"
+        )
+
+    uid = str(uuid.uuid4())
+    file_name = uid + ".wav"
+    # Save the uploaded file
+    with open(file_name, "wb") as buffer:
+        buffer.write(audio_content)
+
+    audio = AudioSegment.from_file(file_name)
+    audio = audio.set_frame_rate(16000)
+
+    audio.export(f"{file_name}", format="wav")
+
+    try:
+        asr_result = asr.audio2text(file_name)
+    except Exception as e:
+        logger.error(e)
+        asr_result = e
+    finally:
+        os.remove(file_name)
+
+    return AudioTranscriptionResponse(text=asr_result)
+
+
 if __name__ == "__main__":
     parser = argparse.ArgumentParser()
     parser.add_argument("--host", type=str, default="0.0.0.0")

comps/asr/whisper/requirements.txt

@@ -7,6 +7,7 @@ opentelemetry-sdk
 prometheus-fastapi-instrumentator
 pydantic==2.7.2
 pydub
+python-multipart
 shortuuid
 transformers
 uvicorn

comps/cores/proto/api_protocol.py

@@ -300,6 +300,38 @@ class AudioChatCompletionRequest(BaseModel):
     user: Optional[str] = None
 
 
+# Pydantic does not support UploadFile directly
+# class AudioTranscriptionRequest(BaseModel):
+#     # Ordered by official OpenAI API documentation
+#     # default values are same with
+#     # https://platform.openai.com/docs/api-reference/audio/createTranscription
+#     file: UploadFile = File(...)
+#     model: Optional[str] = "openai/whisper-small"
+#     language: Optional[str] = "english"
+#     prompt: Optional[str] = None
+#     response_format: Optional[str] = "json"
+#     temperature: Optional[str] = 0
+#     timestamp_granularities: Optional[List] = None
+
+
+class AudioTranscriptionResponse(BaseModel):
+    # Ordered by official OpenAI API documentation
+    # default values are same with
+    # https://platform.openai.com/docs/api-reference/audio/json-object
+    text: str
+
+
+class AudioSpeechRequest(BaseModel):
+    # Ordered by official OpenAI API documentation
+    # default values are same with
+    # https://platform.openai.com/docs/api-reference/audio/createSpeech
+    input: str
+    model: Optional[str] = "microsoft/speecht5_tts"
+    voice: Optional[str] = "default"
+    response_format: Optional[str] = "mp3"
+    speed: Optional[float] = 1.0
+
+
 class ChatMessage(BaseModel):
     role: str
     content: str

comps/tts/gpt-sovits/Dockerfile

@@ -16,7 +16,7 @@ ENV LD_PRELOAD=/usr/lib/x86_64-linux-gnu/libiomp5.so:/usr/lib/x86_64-linux-gnu/l
 ENV MALLOC_CONF="oversize_threshold:1,background_thread:true,metadata_thp:auto,dirty_decay_ms:9000000000,muzzy_decay_ms:9000000000"
 
 # Clone source repo
-RUN git clone https://github.com/RVC-Boss/GPT-SoVITS.git
+RUN git clone --branch openai_compat --single-branch https://github.com/Spycsh/GPT-SoVITS.git
 # Download pre-trained models, and prepare env
 RUN git clone https://huggingface.co/lj1995/GPT-SoVITS pretrained_models
 RUN mv pretrained_models/*  GPT-SoVITS/GPT_SoVITS/pretrained_models/ && \

comps/tts/gpt-sovits/README.md

@@ -54,9 +54,15 @@ wget https://github.com/OpenTalker/SadTalker/blob/main/examples/driven_audio/chi
 
 docker cp chinese_poem1.wav gpt-sovits-service:/home/user/chinese_poem1.wav
 
-http_proxy="" curl localhost:9880/change_refer -d '{
+curl localhost:9880/change_refer -d '{
     "refer_wav_path": "/home/user/chinese_poem1.wav",
     "prompt_text": "窗前明月光，疑是地上霜，举头望明月，低头思故乡。",
     "prompt_language": "zh"
 }'
 ```
+
+- openai protocol compatible request
+
+```bash
+curl localhost:9880/v1/audio/speech -XPOST -d '{"input":"你好呀，你是谁. Hello, who are you?"}' -H 'Content-Type: application/json' --output speech.mp3
+```

comps/tts/speecht5/README.md

@@ -85,8 +85,11 @@ docker run -p 9088:9088 --ipc=host -e http_proxy=$http_proxy -e https_proxy=$htt
 #### 2.2.3 Test
 
 ```bash
-# curl
 curl http://localhost:7055/v1/tts -XPOST -d '{"text": "Who are you?"}' -H 'Content-Type: application/json'
 
+# openai protocol compatible
+# voice can be 'male' or 'default'
+curl http://localhost:7055/v1/audio/speech -XPOST -d '{"input":"Who are you?", "voice": "male"}' -H 'Content-Type: application/json' --output speech.wav
+
 curl http://localhost:9088/v1/audio/speech -XPOST -d '{"text": "Who are you?"}' -H 'Content-Type: application/json'
 ```

comps/tts/speecht5/dependency/speecht5_model.py

@@ -17,33 +17,34 @@ def __init__(self, device="cpu"):
 
             adapt_transformers_to_gaudi()
 
-        model_name_or_path = "microsoft/speecht5_tts"
+        self.model_name_or_path = "microsoft/speecht5_tts"
         vocoder_model_name_or_path = "microsoft/speecht5_hifigan"
-        self.model = SpeechT5ForTextToSpeech.from_pretrained(model_name_or_path).to(device)
+        self.model = SpeechT5ForTextToSpeech.from_pretrained(self.model_name_or_path).to(device)
         self.model.eval()
-        self.processor = SpeechT5Processor.from_pretrained(model_name_or_path, normalize=True)
+        self.processor = SpeechT5Processor.from_pretrained(self.model_name_or_path, normalize=True)
         self.vocoder = SpeechT5HifiGan.from_pretrained(vocoder_model_name_or_path).to(device)
         self.vocoder.eval()
 
         # fetch default speaker embedding
-        if os.path.exists("spk_embed_default.pt"):
-            self.default_speaker_embedding = torch.load("spk_embed_default.pt")
-        else:
-            try:
+        try:
+            for spk_embed in ["spk_embed_default.pt", "spk_embed_male.pt"]:
+                if os.path.exists(spk_embed):
+                    continue
+
                 p = subprocess.Popen(
                     [
                         "curl",
                         "-O",
                         "https://raw.githubusercontent.com/intel/intel-extension-for-transformers/main/"
-                        "intel_extension_for_transformers/neural_chat/assets/speaker_embeddings/"
-                        "spk_embed_default.pt",
+                        "intel_extension_for_transformers/neural_chat/assets/speaker_embeddings/" + spk_embed,
                     ]
                 )
+
                 p.wait()
-                self.default_speaker_embedding = torch.load("spk_embed_default.pt")
-            except Exception as e:
-                print("Warning! Need to prepare speaker_embeddings, will use the backup embedding.")
-                self.default_speaker_embedding = torch.zeros((1, 512))
+            self.default_speaker_embedding = torch.load("spk_embed_default.pt")
+        except Exception as e:
+            print("Warning! Need to prepare speaker_embeddings, will use the backup embedding.")
+            self.default_speaker_embedding = torch.zeros((1, 512))
 
         if self.device == "hpu":
             # do hpu graph warmup with variable inputs
@@ -87,7 +88,9 @@ def _warmup_speecht5_hpu_graph(self):
             "OPEA is an ecosystem orchestration framework to integrate performant GenAI technologies & workflows leading to quicker GenAI adoption and business value."
         )
 
-    def t2s(self, text):
+    def t2s(self, text, voice="default"):
+        if voice == "male":
+            self.default_speaker_embedding = torch.load("spk_embed_male.pt")
         if self.device == "hpu":
             # See https://github.com/huggingface/optimum-habana/pull/824
             from optimum.habana.utils import set_seed

comps/tts/speecht5/dependency/speecht5_server.py

@@ -3,14 +3,21 @@
 
 import argparse
 import base64
+import os
 
 import soundfile as sf
 import uvicorn
 from fastapi import FastAPI, Request
-from fastapi.responses import Response
+from fastapi.responses import Response, StreamingResponse
 from speecht5_model import SpeechT5Model
 from starlette.middleware.cors import CORSMiddleware
 
+from comps import CustomLogger
+from comps.cores.proto.api_protocol import AudioSpeechRequest
+
+logger = CustomLogger("speecht5")
+logflag = os.getenv("LOGFLAG", False)
+
 app = FastAPI()
 tts = None
 
@@ -27,7 +34,7 @@ async def health() -> Response:
 
 @app.post("/v1/tts")
 async def text_to_speech(request: Request):
-    print("SpeechT5 generation begin.")
+    logger.info("SpeechT5 generation begin.")
     request_dict = await request.json()
     text = request_dict.pop("text")
 
@@ -40,6 +47,27 @@ async def text_to_speech(request: Request):
     return {"tts_result": b64_str}
 
 
+@app.post("/v1/audio/speech")
+async def audio_speech(request: AudioSpeechRequest):
+    logger.info("SpeechT5 generation begin.")
+    # validate the request parameters
+    if request.model != tts.model_name_or_path:
+        raise Exception("TTS model mismatch! Currently only support model: microsoft/speecht5_tts")
+    if request.voice not in ["default", "male"] or request.speed != 1.0:
+        logger.warning("Currently parameter 'speed' can only be 1.0 and 'voice' can only be default or male!")
+
+    speech = tts.t2s(request.input, voice=request.voice)
+
+    tmp_path = "tmp.wav"
+    sf.write(tmp_path, speech, samplerate=16000)
+
+    def audio_gen():
+        with open(tmp_path, "rb") as f:
+            yield from f
+
+    return StreamingResponse(audio_gen(), media_type=f"audio/{request.response_format}")
+
+
 if __name__ == "__main__":
     parser = argparse.ArgumentParser()
     parser.add_argument("--host", type=str, default="0.0.0.0")