Замена модели ASR на GigaAM (CTC v2)

app.py

@@ -4,8 +4,8 @@ import subprocess
 import time
 from typing import Dict
 from typing import Optional, Union, List, Tuple
+import gigaam
 
-import whisper
 from fastapi import FastAPI, Depends, HTTPException, UploadFile, File
 from fastapi.security import APIKeyHeader
 
@@ -18,6 +18,7 @@ logger = logging.getLogger(__name__)
 
 app = FastAPI()
 
+model = gigaam.load_model("v2_ctc", device="cuda", download_root="./model")
 
 # API key header
 api_key_header = APIKeyHeader(name="x-api-key")
@@ -108,19 +109,7 @@ def get_audio_duration(file_path: str) -> float:
 async def transcribe_audio(
         file: UploadFile = File(...),
         token: str = Depends(api_key_header),
-        model_name: str = "turbo",
+        model_name: str = "turbo"
-        verbose: Optional[bool] = None,
-        temperature: Union[float, Tuple[float, ...]] = (0.0, 0.2, 0.4, 0.6, 0.8, 1.0),
-        compression_ratio_threshold: Optional[float] = 2.4,
-        logprob_threshold: Optional[float] = -1.0,
-        no_speech_threshold: Optional[float] = 0.6,
-        condition_on_previous_text: bool = True,
-        initial_prompt: Optional[str] = None,
-        word_timestamps: bool = False,
-        prepend_punctuations: str = "\"'\"¿([{-",
-        append_punctuations: str = "\"\'.。,，!！?？:：\")]}、",
-        clip_timestamps: Union[str, List[float]] = "0",
-        hallucination_silence_threshold: Optional[float] = None
 ):
     # Token validation
     if token not in get_keys():
@@ -148,24 +137,29 @@ async def transcribe_audio(
 
         # Transcribe
         logger.info("Starting transcription")
-        result = model.transcribe(
+        if original_duration > 30:
-            temp_output_path,
+            logger.info("Audio duration > 30 seconds, using transcribe_longform")
-            verbose=verbose,
+            transcription_result = model.transcribe_longform(
-            temperature=temperature,
+                temp_output_path
-            compression_ratio_threshold=compression_ratio_threshold,
+            )
-            logprob_threshold=logprob_threshold,
+        else:
-            no_speech_threshold=no_speech_threshold,
+            logger.info("Audio duration <= 30 seconds, using transcribe")
-            condition_on_previous_text=condition_on_previous_text,
+            transcription_result = model.transcribe(
-            initial_prompt=initial_prompt,
+                temp_output_path
-            word_timestamps=word_timestamps,
+            )
-            prepend_punctuations=prepend_punctuations,
+
-            append_punctuations=append_punctuations,
+        full_text = ""
-            clip_timestamps=clip_timestamps,
+        for part in transcription_result:
-            hallucination_silence_threshold=hallucination_silence_threshold
+            if part["transcription"].strip() != "":
-        )
+                full_text += part["transcription"].strip() + " "
+
+        result = {
+            "transcription": transcription_result,
+            "text": full_text
+        }
 
         # Calculate metrics
-        metrics.stop(result["text"], original_duration)
+        metrics.stop(full_text, original_duration)
         logger.info(f"Transcription metrics: {metrics.get_metrics()}")
 
         # Add metrics to result

docker-compose.yml

@@ -1,10 +1,12 @@
+version: '3.8'
+
 services:
   whisper-app:
     build: .
     ports:
       - "9854:9854"
-    devices:
+    volumes:
-      - "/dev/kfd:/dev/kfd"
+      - ./keys.txt:/app/keys.txt
-      - "/dev/dri:/dev/dri"
+      - /tmp:/tmp
-    group_add:
+    command: ["python", "app.py"]
-      - video
+

requirements.txt

@@ -4,3 +4,4 @@ python-multipart
 openai-whisper
 ffmpeg-python
 PyYAML
+numpy<2.0.0