Reducing Jargon Errors with Prompts, Vocabulary Hints, and Post-Processing

Three layers of defense#

Layer 1 — Initial prompt#

whisper.cpp accepts a short text --prompt that acts as prior context for the decoder, nudging it toward expected words without any retraining:

./whisper-cli -m models/ggml-base.en.bin -f standup.wav \
  --prompt "kubectl, Kubernetes, gRPC, staging cluster, CI/CD pipeline, whisper.cpp"

Keep the prompt short and relevant — a few dozen domain terms works better than a huge dumped glossary, since the model treats it as recent conversational context, not a lookup table.

Layer 2 — Choose a bigger or multilingual model for heavy accents#

# small/base models struggle more with strong accents; step up if budget allows
./whisper-cli -m models/ggml-medium.bin -f interview.wav --language auto

--language auto lets whisper.cpp detect the spoken language/accent context rather than forcing an English-only model that penalizes non-native pronunciation patterns.

Layer 3 — Post-processing dictionary correction#

For a fixed, known vocabulary (your own product names, internal tool names), a simple find-and-replace pass on common mis-transcriptions is often more reliable than trying to fix everything via prompting:

CORRECTIONS = {
    "cube cuddle": "kubectl",
    "gee sports": "gRPC",
    "whisper c p p": "whisper.cpp",
}
 
def apply_corrections(text: str) -> str:
    for wrong, right in CORRECTIONS.items():
        text = text.replace(wrong, right)
    return text
 
try:
    with open("raw_transcript.txt") as f:
        raw = f.read()
except FileNotFoundError as err:
    print(f"Error: transcript file missing: {err}")
    raise SystemExit(1)
 
with open("corrected_transcript.txt", "w") as f:
    f.write(apply_corrections(raw))

Build this dictionary from real observed mistakes — run a batch of your actual recordings, diff hypothesis against ground truth, and add every recurring miss.

Error handling for the whole flow#

if [ -z "${1:-}" ]; then
  echo "Usage: $0 <audio-file>" >&2
  exit 1
fi
 
./whisper-cli -m models/ggml-medium.bin -f "$1" --language auto \
  --prompt "$(cat prompts/domain-vocab.txt)" \
  -otxt --output-file raw_transcript \
  || { echo "Error: transcription failed for $1" >&2; exit 1; }

Up next#

The final article looks at multilingual and code-switched speech — speakers who mix languages or accents mid-sentence — and how whisper.cpp's language detection copes with it.