Move environment variables to `config.py`

app/config.py

+import os
+
+import torch
+
+
+class CONFIG:
+    """
+    Configuration class for ASR models.
+    Reads environment variables for runtime configuration, with sensible defaults.
+    """
+    # Determine the ASR engine ('faster_whisper' or 'openai_whisper')
+    ASR_ENGINE = os.getenv("ASR_ENGINE", "openai_whisper")
+
+    # Determine the computation device (GPU or CPU)
+    DEVICE = os.getenv("ASR_DEVICE", "cuda" if torch.cuda.is_available() else "cpu")
+
+    # Model name to use (e.g., "base", "small", etc.)
+    MODEL_NAME = os.getenv("ASR_MODEL", "base")
+
+    # Path to the model directory
+    MODEL_PATH = os.getenv("ASR_MODEL_PATH", os.path.join(os.path.expanduser("~"), ".cache", "whisper"))
+
+    # Model quantization level. Defines the precision for model weights:
+    #   'float32' - 32-bit floating-point precision (higher precision, slower inference)
+    #   'float16' - 16-bit floating-point precision (lower precision, faster inference)
+    #   'int8' - 8-bit integer precision (lowest precision, fastest inference)
+    # Defaults to 'float32' for GPU availability, 'int8' for CPU.
+    MODEL_QUANTIZATION = os.getenv("ASR_QUANTIZATION", "float32" if torch.cuda.is_available() else "int8")
+    if MODEL_QUANTIZATION not in {"float32", "float16", "int8"}:
+        raise ValueError("Invalid MODEL_QUANTIZATION. Choose 'float32', 'float16', or 'int8'.")
+
+    # Idle timeout in seconds. If set to a non-zero value, the model will be unloaded
+    # after being idle for this many seconds. A value of 0 means the model will never be unloaded.
+    MODEL_IDLE_TIMEOUT = int(os.getenv("MODEL_IDLE_TIMEOUT", 0))
+
+    # Default sample rate for audio input. 16 kHz is commonly used in speech-to-text tasks.
+    SAMPLE_RATE = int(os.getenv("SAMPLE_RATE", 16000))

app/faster_whisper/core.py

 import gc
-import os
 import time
 from io import StringIO
 from threading import Lock, Thread
@@ -9,23 +8,20 @@ import torch
 import whisper
 from faster_whisper import WhisperModel
 
-from .utils import ResultWriter, WriteJSON, WriteSRT, WriteTSV, WriteTXT, WriteVTT
+from app.config import CONFIG
+from app.faster_whisper.utils import ResultWriter, WriteJSON, WriteSRT, WriteTSV, WriteTXT, WriteVTT
 
-model_name = os.getenv("ASR_MODEL", "base")
-model_path = os.getenv("ASR_MODEL_PATH", os.path.join(os.path.expanduser("~"), ".cache", "whisper"))
 model = None
 model_lock = Lock()
-
 last_activity_time = time.time()
-idle_timeout = int(os.getenv("IDLE_TIMEOUT", 0))  # default to being disabled
 
 
 def monitor_idleness():
     global model
-    if idle_timeout <= 0: return
+    if CONFIG.MODEL_IDLE_TIMEOUT <= 0: return
     while True:
         time.sleep(15)
-        if time.time() - last_activity_time > idle_timeout:
+        if time.time() - last_activity_time > CONFIG.MODEL_IDLE_TIMEOUT:
             with model_lock:
                 release_model()
                 break
@@ -34,17 +30,11 @@ def monitor_idleness():
 def load_model():
     global model, device, model_quantization
 
-    # More about available quantization levels is here:
-    #   https://opennmt.net/CTranslate2/quantization.html
-    if torch.cuda.is_available():
-        device = "cuda"
-        model_quantization = os.getenv("ASR_QUANTIZATION", "float32")
-    else:
-        device = "cpu"
-        model_quantization = os.getenv("ASR_QUANTIZATION", "int8")
-
     model = WhisperModel(
-        model_size_or_path=model_name, device=device, compute_type=model_quantization, download_root=model_path
+        model_size_or_path=CONFIG.MODEL_NAME,
+        device=CONFIG.DEVICE,
+        compute_type=CONFIG.MODEL_QUANTIZATION,
+        download_root=CONFIG.MODEL_PATH
     )
 
     Thread(target=monitor_idleness, daemon=True).start()

app/openai_whisper/core.py

 import gc
-import os
 import time
 from io import StringIO
 from threading import Lock, Thread
@@ -9,21 +8,19 @@ import torch
 import whisper
 from whisper.utils import ResultWriter, WriteJSON, WriteSRT, WriteTSV, WriteTXT, WriteVTT
 
-model_name = os.getenv("ASR_MODEL", "base")
-model_path = os.getenv("ASR_MODEL_PATH", os.path.join(os.path.expanduser("~"), ".cache", "whisper"))
+from app.config import CONFIG
+
 model = None
 model_lock = Lock()
-
 last_activity_time = time.time()
-idle_timeout = int(os.getenv("IDLE_TIMEOUT", 0))  # default to being disabled
 
 
 def monitor_idleness():
     global model
-    if idle_timeout <= 0: return
+    if CONFIG.MODEL_IDLE_TIMEOUT <= 0: return
     while True:
         time.sleep(15)
-        if time.time() - last_activity_time > idle_timeout:
+        if time.time() - last_activity_time > CONFIG.MODEL_IDLE_TIMEOUT:
             with model_lock:
                 release_model()
                 break
@@ -33,9 +30,15 @@ def load_model():
     global model
 
     if torch.cuda.is_available():
-        model = whisper.load_model(model_name, download_root=model_path).cuda()
+        model = whisper.load_model(
+            name=CONFIG.MODEL_NAME,
+            download_root=CONFIG.MODEL_PATH
+        ).cuda()
     else:
-        model = whisper.load_model(model_name, download_root=model_path)
+        model = whisper.load_model(
+            name=CONFIG.MODEL_NAME,
+            download_root=CONFIG.MODEL_PATH
+        )
 
     Thread(target=monitor_idleness, daemon=True).start()
 

docs/environmental-variables.md

@@ -15,9 +15,11 @@
 export ASR_MODEL=base
 ```
 
-Available ASR_MODELs are `tiny`, `base`, `small`, `medium`, `large`, `large-v1`, `large-v2`, `large-v3`, `turbo`(only OpenAI Whisper) and `large-v3-turbo`(only OpenAI Whisper).
+Available ASR_MODELs are `tiny`, `base`, `small`, `medium`, `large`, `large-v1`, `large-v2`, `large-v3`, `turbo`(only
+OpenAI Whisper) and `large-v3-turbo`(only OpenAI Whisper).
 
-For English-only applications, the `.en` models tend to perform better, especially for the `tiny.en` and `base.en` models. We observed that the difference becomes less significant for the `small.en` and `medium.en` models.
+For English-only applications, the `.en` models tend to perform better, especially for the `tiny.en` and `base.en`
+models. We observed that the difference becomes less significant for the `small.en` and `medium.en` models.
 
 ### Configuring the `Model Path`
 
@@ -28,7 +30,18 @@ export ASR_MODEL_PATH=/data/whisper
 ### Configuring the `Model Unloading Timeout`
 
 ```sh
-export IDLE_TIMEOUT=300
+export MODEL_IDLE_TIMEOUT=300
 ```
 
-Defaults to 0. After no activity for this period (in seconds), unload the model until it is requested again. Setting `0` disables the timeout, keeping the model loaded indefinitely.
+Defaults to `0`. After no activity for this period (in seconds), unload the model until it is requested again. Setting
+`0` disables the timeout, keeping the model loaded indefinitely.
+
+### Configuring the `SAMPLE_RATE`
+
+```sh
+export SAMPLE_RATE=16000
+```
+
+Defaults to `16000`. Default sample rate for audio input. `16 kHz` is commonly used in `speech-to-text` tasks.
+
+