Apple Silicon上的Whisper 2026：Metal基准测试、Core ML设置、M1–M5速度指南

import sounddevice as sd
import numpy as np
from faster_whisper import WhisperModel

model = WhisperModel("large-v3-turbo", device="cpu", compute_type="int8")
buffer = []
chunk_duration = 3
sample_rate = 16000

def callback(indata, frames, time, status):
    buffer.append(indata.copy())
    if len(buffer) * 1024 / sample_rate >= chunk_duration:
        audio = np.concatenate(buffer).flatten().astype(np.float32)
        segments, _ = model.transcribe(audio, beam_size=5)
        for segment in segments:
            print(segment.text)
        buffer.clear()

with sd.InputStream(callback=callback, channels=1, samplerate=sample_rate):
    print("Listening... (Ctrl+C to stop)")
    while True:
        sd.sleep(1000)

import sounddevice as sd
import numpy as np
import requests
import subprocess
from faster_whisper import WhisperModel

WHISPER_MODEL = "large-v3-turbo"
OLLAMA_URL = "http://localhost:11434/api/chat"
LLM_MODEL = "llama3.1:8b"
SAMPLE_RATE = 16000

whisper = WhisperModel(WHISPER_MODEL, device="cpu", compute_type="int8")

def record_audio(duration=5):
    print("Listening...")
    audio = sd.rec(int(duration * SAMPLE_RATE),
                   samplerate=SAMPLE_RATE,
                   channels=1,
                   dtype=np.float32)
    sd.wait()
    return audio.flatten()

def transcribe(audio):
    segments, _ = whisper.transcribe(audio, beam_size=5)
    return " ".join([seg.text for seg in segments])

def llm_respond(user_text):
    response = requests.post(OLLAMA_URL, json={
        "model": LLM_MODEL,
        "messages": [{"role": "user", "content": user_text}],
        "stream": False
    })
    return response.json()["message"]["content"]

def speak(text):
    subprocess.run(
        ["piper", "--model", "en_US-amy-medium.onnx"],
        input=text.encode(),
        check=True
    )

while True:
    audio = record_audio(duration=5)
    user_text = transcribe(audio)
    print(f"You: {user_text}")
    if not user_text.strip():
        continue
    response = llm_respond(user_text)
    print(f"AI: {response}")
    speak(response)