Speech-to-text is usually priced by the amount of audio processed: per second, minute or hour, sometimes converted into credits. The list price per minute is only the starting point. What an organisation actually pays depends on streaming versus batch rates, how each request is rounded, whether stereo channels count twice, which features cost extra, and how much of a monthly commitment goes unused. Self-hosting replaces per-minute fees with GPU and engineering costs. The right comparison is effective cost per hour of your audio, at the accuracy you need.
Key takeaways
- Compare effective cost per hour of your real traffic, not headline per-minute rates.
- Streaming is typically priced higher than batch, because live sessions cannot be batched as efficiently and hold capacity for their full duration.
- Billing increments and per-request minimums can dominate cost for short utterances, such as voice-agent turns or commands.
- Check what is metered: each audio channel, silence and hold time, add-on features and retries can all add to the bill.
- The cheapest transcript is not the cheapest outcome if errors need human correction downstream.
Common pricing models#
| Model | How it works | Suits |
|---|---|---|
| Pay per use | A rate per second, minute or hour of audio, billed monthly | Variable or unpredictable volume |
| Credits or bundles | A plan includes credits; each unit of usage consumes a fixed number | Predictable monthly spend, mixed usage types |
| Committed use | A discounted rate in exchange for a minimum annual or monthly spend | High, stable volume |
| Concurrency or seat-based | A fee per concurrent stream, user or agent, sometimes with unlimited minutes | Contact centres, fixed headcount |
| Licence or self-hosted | A fee for software on your infrastructure, plus your own hardware | On-premise, VPC, strict data control |
Most vendors combine these: a free tier, pay-per-use or credit plans for smaller customers, and negotiated commitments for enterprise volume. The unit is nearly always audio duration, not processing time or transcript length.
What drives the price#
Streaming versus batch#
Streaming sessions tie up server capacity for as long as the audio lasts, must meet latency targets, and batch less efficiently than files. Batch transcription can be scheduled, sorted and packed into large batches on GPUs. Vendors generally reflect this with a higher streaming rate. If audio is already recorded, sending it as a file is almost always cheaper.
Features#
Diarisation, redaction, language identification, summaries, sentiment and domain adaptation may be included, charged per minute on top, or restricted to higher tiers. Enterprise requirements such as on-premise or VPC deployment, service-level agreements and dedicated capacity usually sit in custom contracts.
Model tier and language#
Some vendors price larger or specialised models higher, or charge differently by language. Confirm which model a quoted rate applies to, and that it is the one you evaluated.
Where bills diverge from list prices#
Rounding and minimums#
Each request's duration is rounded up to a billing increment, and some pricing adds a minimum charge per request. For long files this is negligible. For short audio it can multiply the effective rate.
This matters most for voice agents that send each turn as a separate request, and for command-style applications. Streaming a whole call over one session usually avoids it.
Channels, silence and retries#
- Channels. Stereo call recordings with agent and customer on separate channels are often billed per channel, doubling the metered minutes.
- Silence and hold time. Billing is usually on total duration. Trimming long silences with voice activity detection before upload can reduce cost for batch files.
- Retries and reprocessing. Failed or repeated requests, test traffic and re-transcribing an archive after a model change are all billed.
Unused commitment#
Committed plans lower the rate but create a floor. Seasonal traffic, a delayed launch or an over-optimistic forecast can leave paid capacity unused, which raises the effective rate.
Calculating effective cost#
Total spend should include transcription fees, add-on features, unused commitment and the cost of any downstream steps that exist because of transcript errors. The denominator is the audio you needed transcribed, not the audio you were billed for.
The cost of errors#
If transcripts are checked or corrected by people, as in clinical documentation, legal work or captioning, correction time often costs more than the transcription itself. A more accurate system at a higher per-minute rate can be cheaper overall. Measure word error rate and entity accuracy on your own audio as part of the cost comparison. See how to evaluate a speech-to-text vendor.
Self-hosting#
Running an open model yourself turns price into engineering:
A GPU that sustains an RTFx of 100 on your workload processes 100 hours of audio per hour of GPU time at full utilisation. Real utilisation is lower: traffic has peaks and troughs, streaming needs headroom, and capacity must be provisioned for the busiest hour (see concurrency and scaling). Add engineering time, monitoring, security reviews, model updates and on-call support. Self-hosting tends to pay off at high, steady volume, or when data must stay inside your perimeter regardless of cost.
How Reson8 approaches this#
Reson8 uses credits. One minute of real-time streaming costs 2 credits; one minute of batch transcription costs 1 credit. Plans are Free (€0, 600 credits per month, about 5 hours of real-time streaming, one domain model), Pro (€20 per month, 6,000 credits), Growth (€100 per month, 30,000 credits) and Custom, which covers on-premise and VPC deployment and SLAs.
As a worked example, 6,000 Pro credits cover 100 hours of batch audio or 50 hours of streaming, which works out at €0.20 per batch hour or €0.40 per streaming hour when the credits are fully used. Growth gives the same per-hour rate at five times the volume. A team transcribing 40 hours of recorded calls and 20 hours of live streams in a month uses 2,400 + 2,400 = 4,800 credits, which fits within Pro. Current plan terms are in the console.
Common questions#
Why is streaming more expensive than batch?
A live stream occupies capacity for its whole duration, must meet latency targets, and cannot be packed into large, length-sorted batches the way files can. The higher rate reflects lower hardware efficiency.
Is per-second billing always better?
For short requests it is usually cheaper than coarse increments. Compare on your own traffic: for long files, the increment barely matters and the base rate dominates.
Should I self-host to save money?
Only after costing GPUs at realistic utilisation plus engineering and operations time. At modest or spiky volume, a managed API is usually cheaper; at high steady volume or with strict data requirements, self-hosting can make sense.
What should I ask a vendor about pricing?
Billing increment and minimum, how channels are counted, which features cost extra, concurrency limits, what happens when you exceed a plan, and whether the price applies to the model you tested.
References#
- Barroso, L. A., Hölzle, U., & Ranganathan, P. (2018). The Datacenter as a Computer: Designing Warehouse-Scale Machines (3rd ed.). Morgan & Claypool.
- Srivastav, V. et al. Open ASR Leaderboard. Hugging Face.
- Little, J. D. C. (1961). A Proof for the Queuing Formula: L = λW. Operations Research, 9(3).
- FinOps Foundation. FinOps Framework. finops.org.
- Reson8 documentation. docs.reson8.dev.