As checked 2026-10-25: Pre-recorded speech recognition commonly prices the duration of the submitted recording; live recognition can price the duration of an open stream, and text-to-speech may bill input characters or credits. This is a dated economics guide, not a permanent price guarantee or a personal hands-on benchmark.
Scope: global English reference; USD list prices where shown; local eligibility, currency, taxes, negotiated terms and model release dates require separate confirmation. Dynamic source-page prices have a check date, not a verified initial effective date.
Three meters for similar-looking tasks
Pre-recorded speech recognition commonly prices the duration of the submitted recording; live recognition can price the duration of an open stream, and text-to-speech may bill input characters or credits. AssemblyAI’s billing FAQ distinguishes input recording seconds from streaming-session duration. ElevenLabs publishes a shared credit system across speech, music and other products. Deepgram’s Aura-2 speech generation prices characters in blocks of one thousand. These do not translate into each other without separately observed workloads.
Worked per-hour example
A 90-minute file billed at a hypothetical $0.20 per hour costs 1.5 × $0.20 = $0.30, before diarization or other add-ons. A streaming connection left open for two hours could be billed for the session interval rather than the amount of speech captured. The provider’s official billing policy is the controlling source. Do not assume that silence, failure or network reconnection has the same billing treatment in a batch file and a live conversation.
Additional layers
Diarization, translation, timestamps, PII redaction and sentiment or topic analysis may have separate units and prices. A speech-to-speech agent can combine recognition, an LLM, text-to-speech, telephony transport and session orchestration. For production estimates, list each as a separate event stream and record concurrency and peak session duration. A single all-in minutes figure is only appropriate if a provider actually sells such a documented bundle.
Credit reconciliation
If a voice plan grants monthly credits, apply the model’s published credit burn per character or minute, including quality mode where relevant. A credit allowance is a quota, not automatically a fiat-denominated cash balance. Check rollover, expiry, whether API calls spend the same pool, and whether failed generations or regenerations are chargeable. A creative license and the right to clone a voice remain separate questions from the per-minute cost.
Safe decisions
Use prerecorded transcription for material that does not need live captions, and streaming only when timeliness matters. Model data residency, privacy, consent and recording retention before optimizing price. Validate an actual invoice or dashboard event for the target model; the vendor’s current published base rate may exclude bespoke contracts or future promotions.
Audio billed duration is a control surface
An audio ingestion service should read the actual file duration before upload, identify empty or duplicate recordings, and reject files that exceed processing limits. For live sessions, record connected start, connected end and reconnection events in addition to speech-active duration. These clocks may differ by hours when users abandon a browser tab. Reconcile billed session seconds to infrastructure logs each week and create an automatic termination policy consistent with vendor guidance. Incomplete call teardown can turn a modest base tariff into a much larger running charge.
Evaluate privacy and feature add-ons
Speaker identification and diarization are not identical: the first can imply mapping a speaker to a person, while diarization typically partitions voices within a recording. Verify whether the product actually supports the requested feature and separately charges it. Treat consent to record, consent to clone, retention limits and accessibility requirements as independent safeguards. A slightly higher transcription rate may be acceptable for a language that otherwise requires expensive corrections; calculate accepted transcript hours rather than relying on price alone.
Decision checklist
- Identify the specific plan or API endpoint, model version, region and currency before pricing a job.
- Measure input, output, caching, generated media and failed-job units separately.
- Check the official pricing page again before procurement, especially if the SKU has no effective-date notice.
- Keep source evidence and a timestamped price ID; do not treat unknowns as zero.
Pricing evidence boundary
This cost discussion concerns Multi-vendor. Reproduce a quote only with its exact model or plan identifier, chargeable unit, billing region, processing mode and timestamp; a changing vendor rate card is not a permanent price guarantee. If any one of these dimensions is missing, keep the associated rate and worked scenario under editorial review instead of filling the gap with an estimate.
Sources & Further Reading
- AssemblyAI pre-recorded and streaming listed hours — official first-party source, accessed 2026-10-25; original effective date not specified in source.
- AssemblyAI recording duration versus live session duration accounting — official first-party source, accessed 2026-10-25; original effective date not specified in source.
- Deepgram Aura-2 per-character TTS list tariffs and other minutes-based channels — official first-party source, accessed 2026-10-25; original effective date not specified in source.
- ElevenLabs monthly plans, shared credits, credit burn examples and tax exclusion — official first-party source, accessed 2026-10-25; original effective date not specified in source.