mlx-audio STT — Apple Silicon에서 듣기

# Re-verified 2026-08-07 on mlx-audio 0.4.7.
# The old `from mlx_audio.stt.generate import generate` no longer exists;
# the entry point is `generate_transcription`. Two of the parameters were
# renamed with it (`audio_path=` -> `audio=`, `model_path=` -> `model=`);
# `output_path` kept its name but changed meaning — it is now a stem, and
# the extension comes from `format`.
from mlx_audio.stt.generate import generate_transcription

result = generate_transcription(
    model="openai/whisper-tiny",     # see the warning below on WHICH repo
    audio="path/to/audio.wav",
    output_path="transcript",        # extension comes from `format`
    format="txt",                    # txt | srt | vtt | json
)

# `result` is an STTOutput with:
#   .text       — full transcribed text
#   .segments   — list of {id, start, end, text, tokens, ...} per segment
#   .language   — detected language code (e.g. 'en', 'ko')
print(result.text[:500])

import time, wave
from mlx_audio.stt.generate import generate_transcription

audio_path = "path/to/audio.wav"

# Total audio duration via wave module (works for WAV; for MP3/M4A use ffprobe)
with wave.open(audio_path, "rb") as wf:
    duration_sec = wf.getnframes() / float(wf.getframerate())

t0 = time.perf_counter()
result = generate_transcription(model="openai/whisper-tiny", audio=audio_path)
elapsed = time.perf_counter() - t0

print(f"audio duration : {duration_sec:.1f} s")
print(f"transcribe wall : {elapsed:.1f} s")
print(f"realtime factor : {duration_sec / elapsed:.1f}x")

# Measure a REAL file, not a toy. Verifying this lesson on 2026-08-07 I fed it
# a 3.8 s clip and got 0.4x, which says nothing about throughput — model load
# dominated. Load once, transcribe something minutes long, and time only the
# transcribe call if you want a number you can plan capacity with.