Building a Practical ASR Pipeline with whisper.cpp

Overview

A real ASR pipeline is more than one command. It needs: a known-good audio format, a model chosen for your accuracy/speed tradeoff, and a place to put the output.

Step 1 — Normalize the audio

Whisper.cpp expects 16kHz mono WAV. Most source audio isn't in that format, so convert it first with ffmpeg:

ffmpeg -i input.mp4 -ar 16000 -ac 1 -c:a pcm_s16le audio.wav

Step 2 — Pick a model

Model	Size	Relative speed	Use case
`tiny.en`	~75 MB	Fastest	Quick drafts, live captions
`base.en`	~140 MB	Fast	Good default for English
`small.en`	~460 MB	Moderate	Better accuracy, still light
`medium.en`	~1.5 GB	Slow	High accuracy, needs more RAM
`large-v3`	~3 GB	Slowest	Best accuracy, multilingual

bash ./models/download-ggml-model.sh base.en

Step 3 — Transcribe with structured output

./whisper-cli \
  -m models/ggml-base.en.bin \
  -f audio.wav \
  -osrt -otxt -oj \
  --output-file out/audio

This produces out/audio.txt, out/audio.srt, and out/audio.json in one pass.

Step 4 — Wrap it in a script

#!/usr/bin/env bash
set -euo pipefail
 
SRC="$1"
MODEL="${2:-models/ggml-base.en.bin}"
OUT_DIR="out"
 
if [ -z "$SRC" ]; then
  echo "Usage: $0 <input-file> [model-path]" >&2
  exit 1
fi
 
if ! command -v ffmpeg >/dev/null 2>&1; then
  echo "Error: ffmpeg is required but not found on PATH" >&2
  exit 1
fi
 
mkdir -p "$OUT_DIR"
base_name="$(basename "${SRC%.*}")"
wav_path="$OUT_DIR/${base_name}.wav"
 
ffmpeg -y -loglevel error -i "$SRC" -ar 16000 -ac 1 -c:a pcm_s16le "$wav_path" || {
  echo "Error: ffmpeg failed to convert $SRC" >&2
  exit 1
}
 
./whisper-cli -m "$MODEL" -f "$wav_path" -osrt -otxt --output-file "$OUT_DIR/$base_name" || {
  echo "Error: whisper-cli failed to transcribe $wav_path" >&2
  exit 1
}
 
echo "Done: $OUT_DIR/${base_name}.txt"

Common pitfalls

Wrong sample rate: skipping the ffmpeg conversion step gives garbled or empty transcripts.
Model/language mismatch: using an .en-only model on non-English audio produces nonsense text.
Missing disk space check: large-v3 models need several GB free; check before downloading in automated pipelines.

Up next

The final article in this series digs into what happens inside whisper.cpp — the decoder strategies, beam search, and tuning options that affect both speed and accuracy.

Overview#

Step 1 — Normalize the audio#

Step 2 — Pick a model#

Step 3 — Transcribe with structured output#

Step 4 — Wrap it in a script#

Common pitfalls#

Up next#