Adding macos compiled requirements

.gitignore

@@ -158,5 +158,4 @@ cython_debug/
 #  be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore
 #  and can be added to the global gitignore or merged into this file.  For a more nuclear
 #  option (not recommended) you can uncomment the following to ignore the entire idea folder.
-.idea/
+#.idea/
-.vscode/

README.md

@@ -1,22 +1,3 @@
 # LocalWhisper
 
 Making OpenAPI's Open Whisper available via ReST API locally.
-
-## Features
-- [x] Transcription
-- [ ] Translation (Not-planned, but open to PRs)
-
-## Usage Example
-
-Similar to the OpenAI API we can post an audio file to `/audio/transcriptions` as a `multipart/form-data` type.
-
-```
-curl --request POST \
-  --url http://localhost:9000/audio/transcriptions \
-  --header 'Content-Type: multipart/form-data' \
-  --header 'Accept: application/json' \
-  --form 'file=@/path/to/file/example.wav;type=audio/wav'
-  --form 'model=whisper-1'
-```
-
-At present the `model` form field is not required or supported. The `WHISPER_BASE_ASR_MODEL` server configuration will be used to determine the Whisper model to use.

local_whisper/audio.py

@@ -3,9 +3,7 @@ from typing import BinaryIO
 import ffmpeg
 import numpy as np
 
-from .settings import whisper_settings
+DEFAULT_SAMPLE_RATE = 16000
-
-DEFAULT_SAMPLE_RATE = whisper_settings.default_sample_rate
 
 
 # TODO probably can offload this on a worker queue too

local_whisper/inference.py

@@ -4,17 +4,14 @@ from threading import Lock
 import torch
 import whisper
 
-from .settings import whisper_settings
-
 # TODO use pydantic config
-model_name = whisper_settings.base_asr_model
+model_name = os.getenv("ASR_MODEL", "base")
 if torch.cuda.is_available():
     model = whisper.load_model(model_name).cuda()
 else:
     model = whisper.load_model(model_name)
 model_lock = Lock()
 
-
 # TODO move transcribe to a modeling worker
 def transcribe(audio):
     # options_dict = {"task" : task}

local_whisper/settings.py

@@ -1,22 +0,0 @@
-from pydantic import BaseSettings, Field
-
-
-class WhisperSettings(BaseSettings):
-    """Whisper Application Settings.
-
-    All environment varaibles supplied should be prefixed with "WHISPER_".
-    """
-
-    base_asr_model: str = Field(
-        default="medium.en", description="The base whisper model to host."
-    )
-    default_sample_rate: int = Field(
-        default=16000,
-        description="The default sample rate used to resample the audio if necessary",
-    )
-
-    class Config:
-        env_prefix = "WHISPER_"
-
-
-whisper_settings = WhisperSettings()

pyinvoke/autocomplete.zsh

@@ -1,34 +0,0 @@
-# Invoke tab-completion script to be sourced with the Z shell.
-# Known to work on zsh 5.0.x, probably works on later 4.x releases as well (as
-# it uses the older compctl completion system).
-
-_complete_invoke() {
-    # `words` contains the entire command string up til now (including
-    # program name).
-    #
-    # We hand it to Invoke so it can figure out the current context: spit back
-    # core options, task names, the current task's options, or some combo.
-    #
-    # Before doing so, we attempt to tease out any collection flag+arg so we
-    # can ensure it is applied correctly.
-    collection_arg=''
-    if [[ "${words}" =~ "(-c|--collection) [^ ]+" ]]; then
-        collection_arg=$MATCH
-    fi
-    # `reply` is the array of valid completions handed back to `compctl`.
-    # Use ${=...} to force whitespace splitting in expansion of
-    # $collection_arg
-    reply=( $(invoke ${=collection_arg} --complete -- ${words}) )
-}
-
-
-# Tell shell builtin to use the above for completing our given binary name(s).
-# * -K: use given function name to generate completions.
-# * +: specifies 'alternative' completion, where options after the '+' are only
-#   used if the completion from the options before the '+' result in no matches.
-# * -f: when function generates no results, use filenames.
-# * positional args: program names to complete for.
-compctl -K _complete_invoke + -f invoke inv
-
-# vim: set ft=sh :
-

pyinvoke/dev.py

@@ -26,6 +26,6 @@ def build(c):
 
 
 @task
-def serve_dev(c, port=9000):
+def serve_dev(c):
     """Runs the FastAPI webservice"""
-    c.run(f"uvicorn local_whisper.webservice:app --reload --port {port}")
+    c.run("uvicorn local_whisper.webservice:app --reload")

tests/unit/test_inference.py

@@ -1,13 +1,9 @@
-from local_whisper.audio import load_audio
 from local_whisper.inference import transcribe
+from local_whisper.audio import load_audio
 
 
 def test_transcribe(sample_audio):
     with open(sample_audio, mode="rb") as af:
         audio = load_audio(af)
     result = transcribe(audio)
-    assert (
+    assert result["text"].strip() == "Let's see, right now I'm playing Horizon Zero Dawn. I also had just recently finished BioShock Infinite."
-        result["text"].strip().lower()
-        == "Let's see, right now I'm playing Horizon Zero Dawn."
-        " I also had just recently finished BioShock Infinite.".lower()
-    )

tests/unit/test_settings.py

@@ -1,29 +0,0 @@
-import os
-from unittest.mock import patch
-
-from pydantic.types import SecretStr
-
-from local_whisper.settings import WhisperSettings
-
-SETTING_DEFAULTS = {"BASE_ASR_MODEL": "medium.en", "DEFAULT_SAMPLE_RATE": 16000}
-
-
-def test_setting_defaults():
-    """Regression test for settings schema."""
-    with patch.dict(os.environ, {}, clear=True):
-        savant_settings = WhisperSettings()
-        assert len(savant_settings.dict()) == len(SETTING_DEFAULTS)
-        for k, v in SETTING_DEFAULTS.items():
-            _setting_value = getattr(savant_settings, k.lower())
-            unmasked_setting = (
-                _setting_value.get_secret_value()
-                if isinstance(_setting_value, SecretStr)
-                else _setting_value
-            )
-            unmasked_setting == v
-
-
-def test_with_envvar_prefix():
-    with patch.dict(os.environ, {"WHISPER_DEFAULT_SAMPLE_RATE": "22500"}, clear=True):
-        whisper_settings = WhisperSettings()
-        assert whisper_settings.default_sample_rate == 22500