Back to Blog

Choosing Speech & OCR Engines by Cost and Accuracy in 2026

Benchmarks4 min read

There is no useful “best engine” without a source, a task, and a definition of a costly error. A clean single-speaker lecture and a page of handwritten equations ask different things of a model. Choose an engine by the work you need to inspect, then compare the total processing cost.

Read the price units first

The rates below reproduce TomeSift’s R4 provider and economics snapshot, recorded on 19 November 2026. They are wholesale planning inputs for the named batch or model configurations, not TomeSift retail prices, binding provider quotes, or a live price feed. Provider pricing pages remain the place to check current eligibility, minimums, add-ons, and contracts.

An ASR price per audio hour is not an OCR price per thousand pages. Neither includes the whole job: extraction, repair, reasoning, storage, and evidence retrieval can contribute separate costs. Currency also matters: Gladia’s rate is in euros; do not compare it with a dollar rate without an explicit exchange-rate assumption.

Speech: the launch cost snapshot

Wholesale speech processing per audio hour · R4 snapshot, 19 November 2026
EngineRate / hourEvaluation starting point
Groq · Whisper Large v3 Turbo$0.04Fast & Economical; assess speaker labeling separately.
Speechmatics · Batch$0.12Multi-speaker lectures and accents; diarization included in the snapshot.
AssemblyAI · Universal-2 / 3.5 Pro$0.15–$0.23Balanced; check the chosen model. Diarization adds $0.02/hour in the snapshot.
Gladia · Solaria-3 Batch€0.16Multilingual material and automatic language detection.
ElevenLabs · Scribe v2 / v3$0.22Maximum Accuracy candidate; multilingual speech and word timestamps.
Deepgram · Nova-3 Batch$0.26–$0.31Technical vocabulary; distinguish batch from streaming terms.
OpenAI · gpt-transcribe$0.27Domain context and prompting; verify the exact model configuration.

Speaker diarization asks who spoke when; transcription asks what was said. Good words with swapped speaker labels can still make interview notes misleading. Include both in an evaluation if attribution is part of your task. Short recordings can also behave differently under minimum billing rules, so multiply actual billable duration rather than assuming every clip costs a perfectly proportional fraction of an hour.

Documents: layout is part of accuracy

A page can contain correct characters in the wrong reading order. Tables can lose column relationships; equations can lose a superscript or a minus sign. Evaluate structure and source anchors alongside plain text. Select a small set of pages that exposes these failures before sending a whole book.

Wholesale OCR planning inputs per 1,000 pages · same R4 snapshot
EnginePlanning rateWhat to inspect
Docling / MinerU$0 license; ~$0.10–$0.40 computeLayout and bounding-box baseline; self-hosting still costs compute.
Gemini Vision~$0.15–$1.00Messy slides, diagrams, keyframes, and handwriting; token use varies.
Mistral OCR$2.00–$4.00Multilingual academic PDFs; check batch terms.
Reducto$1.00 mini; $10–$15 standardDense tables; model choice changes the cost substantially.
LlamaParse~$1.50+Multi-column documents and reading order; parsing tier matters.
Mathpix$1.50 batch; $5.00 syncDense math, chemistry, physics, and handwriting; image billing differs.

Vision-model page costs are estimates, since image size, page complexity, and output length affect token usage. The snapshot lists eligible batch discounts of 50% for Gemini Vision and Mistral OCR; do not apply a batch discount again to a rate that already includes it. Mathpix’s image rate is a separate unit from its PDF batch rate.

Choose a preset, then inspect a sample

  • Fast & Economical: start with Groq for straightforward audio and a layout extraction baseline such as Docling where configured. Spend review time on unfamiliar terms and structural errors.
  • Balanced: begin with AssemblyAI for speech and a document engine suited to the page layout. Speechmatics is a useful candidate when speaker attribution is central; Mistral OCR is a candidate for academic PDFs.
  • Maximum Accuracy: evaluate ElevenLabs for demanding multilingual audio, and specialist extraction such as Mathpix for dense mathematical material. A higher-priced engine still requires source review.

Presets are configurable starting points. Availability follows the configured service and provider credentials. An engine listed in a guide is not a guarantee that every account can run that model at every moment.

Evaluate without inventing scores

Build a reference from material you are authorized to use: a few representative audio segments, a clean page, a difficult table, and an equation-heavy scan. Manually check the reference against the original. Run each candidate with comparable settings and record the model, date, language, segmentation, and enabled add-ons.

For speech, inspect terminology, negation, numbers, timestamps, and speaker attribution; use word error rate only with a stated normalization method. For OCR, inspect reading order, table cell relationships, equations, and source locations. Keep the raw outputs so cleanup does not hide the extraction errors you intended to measure.

Finally compare usable output per total cost, including human review. TomeSift previews processing credits and reserves the worst-case amount before paid upstream work. Node-level idempotency prevents a retry from becoming a second bill for the same processing node. The point of an economical engine is a useful, reviewable result at a sustainable cost.

Sources & further reading

Provider documentation and research infrastructure behind the discussion. Pricing and terms can change; check the linked provider before choosing a service.