Most Voice AI startups don't fail because of bad code. They fail because their transcription bill scales faster than their revenue. When you're building a meeting assistant, every processed audio hour is a direct line item against your gross margin, and the difference between a healthy margin and a compressed one often traces back to a pricing model you signed up for in hour one. This guide gives you the exact framework to calculate, optimize, and control your transcription cost per audio hour, from your first 500 hours through 50,000 and beyond.
Why speech-to-text costs drive your bottom line
Scaling transcription without margin loss
A meeting assistant running on inefficient audio infrastructure can see gross margin compress significantly, because transcription costs grow with every new active user. And while volume commitments introduce tiered discounts at scale, the cost base expands with your user count in a way that pure software margins do not.
The implication is straightforward: transcription is not a utility cost you optimize once at launch. It's a structural margin driver you need to model at 10x your current volume before committing to any vendor.
Why transcription accuracy shapes downstream cost
Transcription accuracy sets the ceiling for everything downstream. Every CRM entry, coaching score, and AI summary is only as reliable as the words captured in the first layer. When the STT layer misidentifies a speaker, drops a named entity, or garbles a product name, the downstream LLM doesn't recover gracefully, it compounds the error. Teams running higher error rates on accented audio face a different calculation: QA cycles, correction costs, and customer churn when summaries ship with wrong names. Accuracy differences between providers show up most visibly in named entity recognition, speaker attribution, and the quality of every LLM output downstream.
Claap, the video collaboration platform, reached 1-3% Word Error Rate (WER) in production after switching to our infrastructure, with one hour of video transcribed in under 60 seconds. At that accuracy level, AI summaries and action items can flow directly into CRM records without a manual review step: the downstream benefit of getting the transcription layer right.
Unmasking variable cost leaks
Three features make the effective cost per audio hour unpredictable when priced as add-ons:
- Speaker diarization: Highly valuable for attributing words to the correct participant in multi-speaker meetings. For recordings with distinct audio channels per speaker, such as conference calls or podcasts, multichannel transcription is an effective alternative that reduces overlap issues without requiring diarization.
- Translation: Required to serve non-English speakers on global teams
- Sentiment analysis: Surfaces tone and engagement signals in post-meeting summaries and note-taker dashboards. If any of these are billed separately, your variable cost per hour spikes the moment usage grows beyond a single language or meeting format.
Optimizing vendor spend for meeting transcription
The STT pricing landscape
The main managed API providers, Deepgram, AssemblyAI, Speechmatics, and ElevenLabs, all use per-minute billing, but their structures diverge significantly beneath the headline rate. Deepgram charges per minute at a flat rate with most features included at base. AssemblyAI bundles basic diarization into its base rate but gates Speaker Identification, sentiment analysis, entity detection, and summarization behind separate per-minute add-ons. Speechmatics and ElevenLabs follow similar add-on patterns. The result: a headline rate comparison across providers is almost always misleading. Effective cost per audio hour depends on which features your product actually enables, how the billing increment is calculated, and whether volume commitments unlock lower tiers.
Pay-as-you-go versus committed-volume pricing
To make those differences concrete: per Deepgram's public Nova-3 pay-as-you-go pricing ($0.0043/min), a 47-minute meeting costs $0.20. That rate is comparable to our Starter pay-as-you-go tier ($0.61/hr, or roughly $0.48 for the same meeting), while our Growth plan at as low as $0.20/hr requires an upfront volume commitment. The relevant question for modeling is whether your projected monthly volume clears the threshold where a committed-volume contract produces a lower effective rate than pay-as-you-go billing on either side.
Per-hour billing also simplifies financial modeling. Your infrastructure cost per user becomes a direct function of average monthly meeting hours, a metric you already track, rather than a minute-by-minute calculation that obscures true Total Cost of Ownership (TCO).
Identifying hidden billing increment costs
The billing increment trap is one of the most common sources of invoice surprise. Vendors who round up to the nearest 15 seconds or full minute inflate costs significantly on short audio clips and partial streams. A 12-second audio clip billed at a 15-second minimum increases your effective cost by 25% on that clip. The same math applied to a 9-second audio segment billed at a 15-second minimum produces a 67% cost increase on that clip, a figure that compounds across any product handling high volumes of short recordings. Exact-duration billing removes this variable from your cost model entirely.
Starter and Growth plan rates
Our Starter plan runs at $0.61/hr for async transcription. New users receive a one-time €50 in free credits to evaluate production accuracy before committing volume. At 500 hours per month, that's $305 with all audio intelligence features included.
The step-change comes at Growth: as low as $0.20/hr async with upfront commitment, bringing 5,000 hours per month down to $1,000. On Growth and Enterprise plans, your audio is never used to train our models, with no opt-out required.
Mitigating unexpected usage fee spikes
Three operational controls prevent runaway bills:
- Hard usage limits: Configure API-level caps before you're in production, using the concurrency and rate limits documentation as the reference
- Dashboard monitoring: Track consumption in real time to catch spikes before they hit your invoice
- Flat-rate feature packaging: Choose a vendor where diarization, translation, and sentiment don't generate separate line items
We include diarization, translation, sentiment analysis, named entity recognition, and summarization in the base rate on Starter and Growth plans. A spike in multi-speaker meeting volume or translation demand doesn't trigger a separate invoice.
Forecasting speech-to-text bills at 10x growth
The table below models monthly spend at three volume tiers, factoring in all-inclusive pricing versus add-on stacking and DevOps overhead. Self-hosted and legacy figures are estimated from published instance pricing and publicly documented add-on rates.
| Monthly volume |
Gladia (all-inclusive) |
Self-hosted open-source model (AWS G5 + DevOps) |
Legacy cloud / competitors (base + add-ons) |
| 500 hours |
$305 (Starter rate) |
Estimated $900+ (idle GPU + DevOps hours) |
Deepgram Nova-3 pay-as-you-go ($0.0043/min ≈ $0.26/hr, most features included at base rate): ~$130. AssemblyAI (base + Speaker Identification + sentiment + entity detection + summarization at ~$0.30/hr): ~$150. The add-on stack behind the AssemblyAI figure is detailed in the 5,000-hour section below |
| 5,000 hours |
$1,000 ($0.20/hr Growth) |
Infrastructure-dependent. Multi-instance GPU compute plus DevOps engineering hours to manage scaling, VRAM limits, and silence-hallucination failure modes. Costs vary significantly by instance type, utilization rate, and team size |
Deepgram Nova-3 at ~$0.26/hr: ~$1,300. Applying the AssemblyAI effective rate of ~$0.30/hr (base + Speaker Identification + sentiment + entity detection + summarization) to 5,000 hours: ~$1,500. Both figures are before any volume discount that may apply |
| 50,000 hours |
Custom Enterprise |
Infrastructure-dependent. Multi-GPU cluster provisioned for peak concurrency, dedicated DevOps coverage for VRAM management, silence-hallucination monitoring, and model versioning. Costs vary significantly by instance type, reserved vs. on-demand pricing, and team headcount |
Deepgram Nova-3 at ~$0.26/hr: ~$13,000, before any volume discount that may apply at this tier. Applying the AssemblyAI effective rate of ~$0.30/hr to 50,000 hours: ~$15,000, before any additional per-feature charges your product requires, and before any volume discount that may apply at this tier |
Modeling transcription spend at 500 hours
At 500 hours per month, the Starter rate of $0.61/hr lands at $305 with all audio intelligence features included: diarization, translation, sentiment analysis, named entity recognition, and summarization. The self-hosted cost model at this utilization profile, including idle GPU compute and DevOps overhead, is covered in the build vs. buy section below.
Cost projections for 5,000 audio hours
At 5,000 hours per month, the all-inclusive Growth plan at $0.20/hr lands at $1,000 flat. Deepgram Nova-3 at its public pay-as-you-go rate of $0.0043/min (≈ $0.26/hr) comes to approximately $1,300, with most features included at base, a lower effective rate than AssemblyAI but higher than our committed-volume Growth tier. Per AssemblyAI's public pricing at time of writing, their Universal-2 base rate of $0.15/hr includes basic speaker diarization, but add-ons for Speaker Identification ($0.02/hr), sentiment ($0.02/hr), entity detection ($0.08/hr), and summarization ($0.03/hr) push the effective rate to approximately $0.30/hr, or roughly $1,500 at this volume. The gap between all-inclusive and add-on pricing widens with every feature your product enables.
Modeling TCO at 50,000 hours
At Series A/B scale, self-hosting requires a multi-GPU cluster with dedicated DevOps coverage. Our Enterprise plan is custom-priced and includes zero data retention and dedicated cloud clusters in EU and US regions for compliance-driven teams. Aircall's trajectory is the most direct reference point at scale: they cut transcription time by 95% (from 30 minutes to 1.5 minutes per call) and now process over 1 million calls per week through our infrastructure, with no capacity pre-provisioning.
Negotiating custom volume discounts
When you approach Enterprise negotiations, the four levers to prioritize are: debundled pricing for features you don't need, custom model fine-tuning for your specific audio environment, SLA guarantees tied to your product's uptime requirements, and zero data retention guarantees for compliance. Teams processing more than 10,000 hours per month should open this conversation before renewing at Growth rates.
How peak concurrency inflates your hourly spend
Meeting starts concentrate at predictable times each week. Self-hosted infrastructure must be sized for that peak, leaving GPUs at 10-15% utilization during off-peak hours. You're paying for peak capacity 24 hours a day, 7 days a week, to handle usage that concentrates for 2-3 hours daily.
Our infrastructure handles thousands of parallel calls with no pre-provisioning and no capacity forecasting on your side. Meeting-assistant traffic follows predictable weekly patterns (heavy Monday mornings, lighter Fridays) and our infrastructure absorbs those spikes without any capacity pre-provisioning on your side.
Root causes of infrastructure bill bloat
Eliminating duplicate API call overhead
Retries on failed webhooks, broken WebSocket reconnections, and duplicate polling logic can cause you to transcribe the same audio segment twice. At $0.20/hr, a 1-hour recording transcribed twice costs $0.40 instead of $0.20. Across thousands of daily recordings, deduplication at the application layer is worth the engineering investment. The API supports idempotency patterns that prevent duplicate submissions at the request level.
Eliminating dormant feature billing
Our model on Starter and Growth is simple: every audio intelligence feature is included in the base rate, so you can configure only the outputs your product actually uses (diarization, translation, or entities) without paying for the rest.
Reducing pre-transcription file overhead
Uncompressed WAV at 44.1 kHz stereo, 16-bit, runs to approximately 600 MB per hour of audio, a figure derivable from the sample rate and bit depth (44,100 samples/sec × 2 channels × 2 bytes = ~176 KB/sec). Compressed AAC or M4A at 128–192 kbps bitrate runs to approximately 60–90 MB per hour, depending on encoder settings. That 6-10x difference in file size affects upload time, storage costs, and egress fees. We accept WAV, M4A, FLAC, and AAC natively, so there's no format conversion required to ship compressed audio.
Raw audio storage on AWS S3 Standard costs $0.023/GB per month at time of writing, and egress to the internet runs $0.09/GB for the first 10 TB/month. Implementing a strict retention policy, deleting raw audio after the transcript is confirmed and storing only the structured text output, cuts storage costs significantly and removes egress fees from your recurring bill. The delete transcription API endpoint makes automated post-processing cleanup straightforward.
Build vs. buy: self-hosted model cost overhead
Self-hosted model TCO: GPU and compute overhead
Self-hosting an open-source model on AWS g5.xlarge on-demand costs $1.006/hr at time of writing, or $734/month for a single instance running continuously. At 500 hours of monthly audio volume with burst-pattern processing typical for meeting assistants, GPU utilization can run quite low, meaning you're paying for substantial idle compute. The effective cost per processed audio hour on a self-hosted open-source model at low utilization rates can exceed our Starter plan rate of $0.61/hr significantly once you account for idle time and engineering overhead.
Production deployments of open-source models also carry failure modes that don't appear in demos: Video Random Access Memory (VRAM) exhaustion on large files, hallucination of repeated text during silence periods, and timestamp drift on compressed audio. Each of these becomes an unplanned debugging session.
Hidden dev overhead in self-hosting
Every sprint point your team spends debugging GPU scaling is a sprint point not spent on the AI summary quality, CRM integration, or calendar sync that is your actual 10x differentiator. Teams moving off self-hosted open-source models to a managed API report saving more than 20% in DevOps effort, capital that goes directly into product velocity.
The "self-hosting is free" assumption breaks the moment you account for GPU instance costs, idle compute, and your engineers' hourly rate. Production-grade speaker attribution alone adds VRAM management, speaker embedding alignment, and model versioning to your backlog. It's only free if you ignore the inputs.
Architecture decisions that reduce cost per hour
Choosing the right transcription model for your audio
Model selection directly affects both accuracy and cost efficiency. For European business audio (English, French, German, Spanish, Italian), Solaria-3 ranks #1 on real customer recordings against AssemblyAI, ElevenLabs, Deepgram, and Mistral. On Switchboard conversational data it posts 33.9% WER, the only model under 35%. Higher accuracy at the transcription layer means fewer correction cycles and fewer hallucinations in your downstream LLM outputs, the layer that defines meeting assistant product quality.
For global products requiring maximum language breadth, code-switching, or real-time streaming, Solaria-1 covers 100+ supported languages.
| Model / provider |
Production accuracy |
European languages support |
Downstream impact |
| Solaria-3 (Gladia) |
Ranked #1 on real customer recordings across EN, FR, DE, ES, IT |
Yes (optimized EN, FR, DE, ES, IT) |
Built for noisy, conversational, and accented business audio across EN, FR, DE, ES, IT |
| Solaria-1 (Gladia) |
Strong multilingual accuracy across 100+ languages |
Yes (broad language coverage, code-switching) |
Excellent for global multilingual products |
| Competitor Nova-3 |
Competitive on varied audio types |
Multilingual support available, including select European languages |
Good performance on varied audio types |
Batch vs. streaming cost profiles
Async transcription is not only cheaper (as low as $0.20/hr vs. as low as $0.25/hr on Growth, depending on committed volume) but also more accurate because the model analyzes the full recording context before producing the final transcript. For meeting assistants, where notes are generated post-meeting rather than during the call, async is the right default on both counts.
Speaker diarization powered by pyannoteAI's Precision-2 model is available in async workflows only. Per pyannoteAI's published benchmarks, Precision-2 is 14% more accurate than its predecessor and 28% more accurate than the pyannote.audio OSS 3.1 model, requiring full-batch processing for speaker embedding alignment across the entire conversation. If your product needs accurate speaker attribution, that's another reason async is the correct architecture.
Why real-time transcription costs more than async
Real-time transcription requires persistent WebSocket connections and dedicated compute, which is why our real-time rate (as low as $0.25/hr on Growth) is higher than our async rate (as low as $0.20/hr). For meeting assistants generating post-meeting notes and summaries, async is the right choice on both cost and accuracy grounds. Our async pipeline processes approximately one hour of audio in under 60 seconds, giving users notes within seconds of a meeting ending. For meeting assistant products, the real-time vs async tradeoff almost always resolves in favor of async.
The hidden cost table below captures the secondary costs that don't appear on a vendor's pricing page but accumulate across a production deployment.
| Cost category |
Standard industry rate |
Impact on meeting assistant SaaS |
| Cloud storage |
$0.023 per GB per month |
Storing raw WAV files bloats monthly hosting bills fast |
| Data egress |
$0.09 per GB (first 10 TB/month), decreasing to $0.085/GB up to 50 TB, $0.07/GB up to 150 TB, and $0.05/GB beyond that |
Moving large audio files between servers incurs transfer fees |
| DevOps maintenance |
$60–120 per engineer hour (senior DevOps/architect market rates) |
Debugging GPU scaling, silence hallucinations, and file limits burns runway |
Reducing transcription spend per hour
The compounded playbook:
- Compress audio to AAC or M4A before submission to cut file sizes by 6-10x
- Route all meeting workflows through async rather than streaming to save $0.05/hr and improve accuracy
- Choose all-inclusive pricing at the Growth tier so diarization and translation spikes don't generate add-on charges
- Implement deduplication at the application layer to prevent double-billing on retries
- Transition to committed-volume pricing before you hit 1,000 hours per month. The blind provider comparison tool lets you run your own audio against multiple providers rather than relying on any vendor's benchmark.
"Accurate, fast and developer friendly transcription API for multilingual audio" - Faes W. on G2
Start with €50 in free credits and have your integration in production in less than a day. Our standard REST and WebSocket protocols mean you're not locked in.
FAQs
What is a competitive transcription cost per audio hour for a meeting assistant?
For teams processing 500-5,000 hours per month with diarization, translation, and sentiment included, a competitive all-in rate falls between as low as $0.20/hr (Growth, volume-committed) and $0.61/hr (Starter, pay-as-you-go) on a managed API. A self-hosted open-source model on AWS g5.xlarge at typical early-stage utilization levels can produce significantly higher effective costs per processed audio hour when accounting for idle GPU time and DevOps overhead.
How do I control my monthly transcription spend as my user base grows?
The three highest-impact controls are: moving from pay-as-you-go Starter pricing to a committed-volume Growth plan once you pass 1,000 hours per month, choosing a vendor with flat all-inclusive feature pricing so diarization and translation spikes don't generate add-on charges, and compressing audio to AAC or M4A before submission to reduce file sizes and eliminate egress fees between storage and the transcription API.
At what volume should I move to a custom enterprise contract?
Teams processing substantial monthly volumes should open Enterprise conversations to access debundled pricing, custom model fine-tuning, dedicated Service Level Agreements (SLAs), and zero data retention guarantees. Our compliance hub covers the SOC 2 Type II, ISO 27001, HIPAA, HDS, and GDPR documentation Enterprise procurement teams typically request.
Does Gladia use my audio data to train its models?
On Growth and Enterprise plans, your audio is never used for model training, and no opt-out action is required. On the Starter plan, data can be used for model training by default, so if data privacy is a hard requirement from an enterprise customer's legal team, Growth is the minimum tier that provides this guarantee.
What is the accuracy difference between Solaria-3 and Solaria-1 for European business audio?
Solaria-3 ranks #1 against AssemblyAI, ElevenLabs, Deepgram, and Mistral on real customer recordings across English and core European languages, with a measurable accuracy advantage over Solaria-1 on noisy, conversational, and accented business audio. Solaria-1 remains the better choice for products needing code-switching support across a broad range of languages or real-time streaming.
Key terms glossary
Word error rate (WER): The percentage of words a transcription model gets wrong, counting substitutions, insertions, and deletions, compared to a verified reference transcript. Lower is better.
Diarization error rate (DER): The percentage of a transcript where speaker attribution is wrong, factoring in missed speech, false alarms, and confused speaker labels.
Diarization: The process of labeling which speaker said what in a multi-speaker recording. Diarization is powered by pyannoteAI's Precision-2 model and is available in async workflows only.
Code-switching: When a speaker changes languages mid-conversation. A model handles code-switching correctly if it keeps up with the language change without breaking the session or garbling the output.
Data egress: The fee cloud providers charge when data leaves their network, such as downloading audio files or transcripts. Storing less raw audio for a shorter period reduces this cost.
Idempotency: A property of an API request where making the same call multiple times produces the same result as making it once, preventing accidental duplicate processing, and duplicate billing, from retries.
REST: The standard web API pattern for one-off request and response calls, used here for async transcription submissions.
