distil-whisper

# distil-whisper via Hugging Face transformers >=4.39
pip install --upgrade transformers accelerate datasets[audio]

# transcribe.py
import torch
from transformers import pipeline

device = "cuda:0" if torch.cuda.is_available() else "cpu"
dtype  = torch.float16 if torch.cuda.is_available() else torch.float32

pipe = pipeline(
    "automatic-speech-recognition",
    model="distil-whisper/distil-large-v3.5",  # or distil-large-v3, distil-small.en
    torch_dtype=dtype,
    device=device,
    chunk_length_s=25,   # long-form: chunk into 25s windows
    batch_size=16,       # GPU only; on CPU set batch_size=1
)

result = pipe("meeting.wav", return_timestamps=True)
print(result["text"])
for chunk in result["chunks"]:
    print(chunk["timestamp"], chunk["text"])

# CTranslate2 path · production sweet spot
pip install faster-whisper

from faster_whisper import WhisperModel, BatchedInferencePipeline

model = WhisperModel(
    "Systran/faster-distil-whisper-large-v3",
    device="cuda",
    compute_type="float16",   # int8_float16 halves VRAM
)

# Parallel decode across 30s chunks
batched = BatchedInferencePipeline(model=model)

segments, info = batched.transcribe(
    "podcast.wav",
    batch_size=16,            # 8 on 8 GB, 16 on 16 GB, 32 on 24 GB+
    beam_size=5,
    word_timestamps=True,
)
for s in segments:
    print(f"[{s.start:6.2f} -> {s.end:6.2f}] {s.text}")

# whisper.cpp + distil-large-v3 GGML (laptop / edge)
git clone https://github.com/ggml-org/whisper.cpp
cd whisper.cpp
make

# fetch the official GGML conversion (auto-grabs ggml-distil-large-v3.bin)
bash ./models/download-ggml-model.sh distil-large-v3

# transcribe a 16kHz mono wav (resample first if needed)
./build/bin/whisper-cli \
  -m models/ggml-distil-large-v3.bin \
  -f samples/meeting.wav \
  -l en \
  --output-srt --output-vtt

# quantise to ~half the disk + RAM with a small WER cost
./build/bin/quantize \
  models/ggml-distil-large-v3.bin \
  models/ggml-distil-large-v3-q5_0.bin q5_0