Deepgram → faster-whisper
What it saves, what actually moves, what you rebuild — and the thing that catches people.
What it costs, and what it saves
From roughly $0.0043 a minute to nothing per minute. A product transcribing 1,000 hours a month goes from about $260 to the cost of the electricity, and a one-off archive ingest stops being a four-figure invoice entirely.
Moves cleanly
Everything, because transcription is stateless. Audio in, text out — there is no accumulated state, no schema and no history to port. Your existing audio files and any transcripts you have already stored are untouched.
You rebuild
The response parsing. Deepgram returns its own JSON shape with word objects and speaker labels; faster-whisper returns segments with optional word timestamps. Expect a day rewriting whatever consumes the transcript, and note that diarization is not included — pair with WhisperX if you need speaker labels.
What Deepgram costs you today
Pay-as-you-go from roughly $0.0043 per minute on the standard model, with volume discounts and enterprise contracts above that. A free credit grant covers early development. It reads as trivially cheap until you multiply: a product transcribing a thousand hours a month is in four-figure territory, and transcription volume tends to grow with exactly the usage you were hoping for.
What actually holds you in
Very low — the lowest in this batch. Transcription is stateless: audio in, text out. There is no accumulated state to migrate and no schema to rebuild. The integration is one API call, and several of the alternatives below expose a compatible server interface. What differs is output formatting and the shape of diarization results, so budget a day of parsing adjustments rather than a project.
What you are moving to
faster-whisper reimplements Whisper inference on CTranslate2, delivering roughly a four-fold speedup over the reference implementation with substantially lower memory use, and the same transcription output — these are the same weights, executed better. It supports int8 and float16 quantisation, batching, and word-level timestamps, and runs on both GPU and CPU. For most teams replacing a paid transcription API, this is simply the correct starting point.
Free. Hardware you already own; a laptop handles the smaller models.
faster-whisper strengths
- Several times faster than reference Whisper at equal accuracy
- Quantisation options let large models fit modest GPUs
- Runs on CPU when no GPU is available
- MIT licensed, no per-minute cost, nothing leaves your machine
What you give up
- Batch-oriented; streaming needs extra work to do well
- Diarization is not included — pair with WhisperX or pyannote
- Accuracy varies by language more than the managed services do
The migration, step by step
- 1Benchmark first, on your own audio. Take twenty representative files — the hardest ones, not the cleanest — and run them through both. Accuracy is very domain-dependent and a public benchmark will not tell you what yours does.
- 2Install faster-whisper and pull a model. `pip install faster-whisper`, then load `large-v3` if you have an 8GB+ GPU, or `medium` on CPU. Quantize to int8 if memory is tight.
- 3Wrap it in the same interface your code already calls. Keep the Deepgram-shaped function signature and change only its body — that keeps the blast radius to one file.
- 4Run both in parallel for a week, writing both transcripts to storage. Compare on real traffic rather than on your benchmark set.
- 5Move batch and archive work over first. That is where per-minute pricing hurts most and where latency does not matter at all.
- 6Keep Deepgram for live streaming if you have it, and cut over only the batch path. A partial migration is a legitimate end state here.
The gotcha
Whisper hallucinates on silence. Long pauses, music, or dead air can produce confident invented sentences — often repeated phrases from the training distribution. Deepgram does not do this. Set a voice-activity-detection filter (faster-whisper ships one) and clamp `no_speech_threshold`, or you will ship transcripts containing text nobody said.
When to stay on Deepgram
You need real-time streaming with sub-second interim results — that is the one place the managed services still lead clearly and reproducing it is a project, not a migration. Stay too if your volume is a few hours a month, where the bill is genuinely smaller than running a GPU, or if you need consistent quality across many less-common languages and have no appetite for evaluating models per language.
More Layer 4 migrations
Figures verified 2026-08-11 against vendor pricing pages. Prices change and migrations differ by estate — treat the cost delta as a starting model, not a quote. Rankings and recommendations here are merit-only; affiliate income never changes a verdict. See our methodology.