insanely-fast-whisper

# insanely-fast-whisper 0.0.15 · Python >=3.8
pipx install insanely-fast-whisper

# Transcribe a local file or a URL.
# Writes output.json (chunked timestamps + text) in cwd.
insanely-fast-whisper --file-name input.mp3

# Pin the version inside CI / scripts:
# pipx install insanely-fast-whisper==0.0.15 --force

# Python 3.11.x and pipx complains about requires-python? Use:
# pipx install insanely-fast-whisper --force \
#   --pip-args="--ignore-requires-python"

# Step 1 — build flash-attn against your CUDA stack.
# Ampere (A100), Ada (RTX 4090), or Hopper (H100) only.
pip install flash-attn --no-build-isolation

# Step 2 — install the CLI itself.
pipx install insanely-fast-whisper

# Step 3 — run with Flash Attention 2 + batched chunks.
insanely-fast-whisper \
  --file-name podcast.mp3 \
  --model-name openai/whisper-large-v3 \
  --batch-size 24 \
  --flash True \
  --timestamp chunk \
  --transcript-path output.json

# distil-whisper variant — ~20% faster again, English-focused:
# --model-name distil-whisper/distil-large-v2

# pip install transformers torch pyannote.audio
import torch
from transformers import pipeline
from pyannote.audio import Pipeline as Diarizer

asr = pipeline(
    "automatic-speech-recognition",
    model="openai/whisper-large-v3",
    torch_dtype=torch.float16,
    device="cuda:0",
    model_kwargs={"attn_implementation": "flash_attention_2"},
)

out = asr(
    "meeting.wav",
    chunk_length_s=30,
    batch_size=24,
    return_timestamps=True,
)

# pyannote needs a HF token + accepted gated model conditions.
diar = Diarizer.from_pretrained(
    "pyannote/speaker-diarization-3.1",
    use_auth_token="<HF_TOKEN>",
).to(torch.device("cuda"))

speakers = diar("meeting.wav")
# Merge ASR chunks with speaker turns by overlap on the timeline.