API Comparison Table

Heading 1

Heading 2

Heading 3

Heading 4

Heading 5
Heading 6

Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua. Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris nisi ut aliquip ex ea commodo consequat. Duis aute irure dolor in reprehenderit in voluptate velit esse cillum dolore eu fugiat nulla pariatur.

Block quote

Ordered list

  1. Item 1
  2. Item 2
  3. Item 3

Unordered list

Text link

Bold text

Emphasis

Superscript

Subscript

Pricing
Get started
Get started

Read more

Product News

Gladia to join OVH Groupe, accelerating Europe's sovereign voice AI ambitions

PARIS, France — Sept 21, 2026 — Gladia, the AI audio infrastructure provider trusted by more than 300,000 developers and 2,000 enterprise customers worldwide, announced July 31st, that it has been acquired by OVH Groupe, the parent company of OVHcloud and Europe's leading sovereign cloud provider.

Speech-To-Text

Migrating from Azure Speech to Gladia: a step-by-step switching guide

TL;DR: Migrating from Azure Speech to Gladia removes the overhead of custom training pipelines and fragmented per-feature billing. Azure routes diarization, translation, and sentiment through separate services with separate billing meters. We bundle all audio intelligence into one per-hour rate on Starter and Growth plans. Solaria-3 ranks #1 for real-world European business audio, Solaria-1 covers 100+ languages with native code-switching. Both deliver out-of-the-box accuracy that eliminates custom training for most production audio. Most engineering teams complete the API refactoring in under 24 hours.

Speech-To-Text

How to evaluate a speech-to-text API: a technical buyer's framework

TL;DR: Choosing an STT API on vendor benchmarks alone is how teams end up with transcription that looks fine in staging and breaks on production audio. A rigorous evaluation requires building a test set from your own calls, measuring word error rate (WER) on your specific audio distribution, stress-testing latency under concurrent load, and auditing data retraining terms before signing. This guide gives you a reusable engineering blueprint to run that evaluation end-to-end, the same methodology behind our own open async benchmark, which covers 7 datasets and 74+ hours of audio across 8 providers.

Enterprise and on-premise speech-to-text: security, SLA, and data residency

Published on August 28, 2026
by Ani Ghazaryan
Enterprise and on-premise speech-to-text: security, SLA, and data residency

TL;DR: If your speech-to-text vendor trains their models on your customers' audio by default, you're carrying a silent compliance liability into every product deployment. Building enterprise-ready voice products demands more than low Word Error Rates: it requires absolute control over data residency, strict compliance with frameworks like GDPR and SOC 2 Type II, and predictable cost models that don't collapse at scale. This guide covers what to verify before committing to an enterprise STT vendor, including whether you genuinely need on-premise infrastructure or whether cloud-native controls (ZDR, EU data residency, per-account isolation, DPA/BAA) already satisfy the same governance requirement, and explains exactly where we draw the line on data handling across every pricing tier. Gravite, a French call quality-monitoring platform transcribing 50,000 hours of audio per year, cut call quality review time by 93% on our cloud API, without managing a single piece of on-premise infrastructure.

If you're evaluating speech infrastructure for a high-volume voice product, the base transcription rate is rarely the number that breaks your unit economics. What breaks them is the compounding cost of add-on features, the legal exposure from a vendor's default data-training policy, and the engineering overhead of patching together three separate providers for recording, transcription, and enrichment. This article addresses the compliance, security, SLA, and deployment objections that surface late in every enterprise evaluation cycle.

Mitigating risk in production speech deployments

Every CRM entry, coaching score, and AI summary in a voice product is ceiling-bounded by the quality of the transcript that feeds it. A wrong name silently corrupts a sales record. A missed consent statement creates a compliance exposure. A garbled account number in a contact center interaction produces a downstream error the QA team won't catch for weeks. These aren't transcription problems in isolation, they're pipeline-wide failures that originate at the first layer and compound downstream.

Most teams build their audio pipeline with separate providers: one for recording, one for transcription, one for enrichment. Each integration point degrades data quality, drops metadata, or creates an SLA gap between vendors. Consolidating to a single audio infrastructure layer removes those seams and makes the failure surface predictable.

Establishing data governance boundaries and navigating privacy laws

Data governance in a speech pipeline starts with mapping exactly where audio travels: from raw capture through transcription, enrichment, storage, and deletion. For regulated industries, "the vendor processes it in the cloud" is not a sufficient answer. The governance questions to resolve before signing any enterprise STT contract:

  • Which cloud regions process and store audio, and are EU workloads kept within EU boundaries by default?
  • What is the default data retention period, and can it be set to zero?
  • Does the vendor use customer audio to retrain models by default, and on which pricing tiers?
  • Are a Data Processing Agreement (DPA) or Business Associate Agreement (BAA) available for your jurisdiction?

GDPR Article 32 requires "appropriate technical and organisational measures" to protect personal data, which includes voice recordings. For EU-based products, audio processed outside the EU without an adequacy decision or Standard Contractual Clauses creates legal exposure. HIPAA adds a separate layer for healthcare: audio containing Protected Health Information requires a signed BAA with the processing vendor. The data residency requirements for voice data differ meaningfully between US and EU regulatory contexts, and vendors that treat residency as an enterprise add-on rather than a default configuration create friction at exactly the wrong moment in a compliance review.

Enterprise speech-to-text cost models

Most product leaders model their STT costs on the base transcription rate, then discover that diarization, named entity recognition (NER), sentiment analysis, and translation are each metered separately. At 10,000 hours per month, those add-ons can substantially increase the effective per-hour rate. The gap compounds quickly when diarization, NER, sentiment, and translation are each metered separately.

Our pricing on Starter and Growth plans bundles every audio intelligence feature at the base rate with no per-feature surcharges. The table below shows predictable scaling across monthly volumes, with all features (diarization, NER, sentiment, translation, custom vocabulary) included.

Table 1: Total cost of ownership at scale (all features included)

Monthly volume Starter plan Growth plan Enterprise plan
100 hours $61 async / $75 real-time From $20 async Custom
1,000 hours $610 async / $750 real-time From $200 async Custom
10,000 hours $6,100 async / $7,500 real-time From $2,000 async Custom

Growth plan rates reflect volume commitment discounts. Enterprise includes fine-tuning, custom models, and custom SLAs. All audio intelligence features included on Starter and Growth with no add-on fees.

By contrast, per Deepgram's public pricing, Nova-3 Multilingual real-time streaming charges approximately $0.55/hr, with diarization billed as a separate streaming add-on ($0.0020/min) on top of the base rate. At 10,000 hours per month with a full audio intelligence payload enabled, the gap between bundled and add-on pricing structures becomes material in the annual plan model.

Cloud vs. on-premise speech-to-text: the real trade-offs

The cloud-versus-on-premise decision is rarely binary for enterprise teams. The answer depends on your compliance requirements, data sensitivity, engineering capacity, and the audio volumes you're processing across different product lines.

Table 2: Cloud API vs. self-hosted STT: an honest comparison

Concern Cloud API (Gladia) Self-hosted / on-premise
Data residency Configurable by region (EU, US) Fully within your VPC or data center. Gladia does not offer on-premise or air-gapped hosting on any plan. If literal on-premise is a hard requirement, a self-hosted open-source model or a vendor that explicitly supports on-prem is the correct path.
Real-time latency ~300ms final latency; partial transcripts under 103ms. Applies to Solaria-1 real-time streaming. Determined entirely by your hardware and model. Self-hosted models on commodity GPUs typically add significant inference overhead compared to managed cloud infrastructure. No vendor SLA applies.
Async processing throughput Approximately 60 seconds to process one hour of audio. A throughput ratio, not a per-utterance latency figure. Applies to batch (async) workflows. Depends on GPU hardware, model size, and batch configuration. Throughput is fully customer-managed with no vendor benchmark or SLA.
Maintenance overhead Zero (vendor-managed) Fully customer-managed: your team handles model deployment, upgrades, scaling, and incident response with no vendor SLA or support.
Cost predictability Per-hour, no add-ons on paid plans Hardware procurement plus staffing costs vary significantly based on concurrency, GPU requirements, and team size. Total cost of ownership is typically underestimated when infrastructure, scaling, and ongoing maintenance are fully accounted for. No vendor pricing applies.
Mixed workload operations Vendor-managed Not applicable. Gladia is cloud-only. Mixed workloads that require literal on-premise components must use a separate self-hosted model for those segments. Cloud-routed audio can use the Gladia API in parallel.

Operational benefits of cloud STT

Cloud APIs eliminate infrastructure provisioning entirely. Most product teams reach production in under 24 hours using our lightweight Python or JavaScript SDKs, with no hardware to pre-configure and no capacity to forecast. Aircall cut transcription time by 95% (from 30 minutes to 1.5 minutes per call) and now processes over 1 million calls per week through our API, scaling to handle demand spikes without pre-provisioning additional compute. When a fintech customer needs to run 800 concurrent sessions, the infrastructure spins up in seconds with no manual intervention.

When on-premise is actually required, and when it isn't

The phrase "data must never leave our environment" covers two genuinely different requirements, and confusing them leads to overbuying infrastructure you don't need.

  • True air-gap / literal on-premise: Defense contractors operating under classified network requirements, intelligence agencies, and sovereign financial institutions with physical isolation mandates need audio to be processed on hardware that has no external internet connection under any circumstances. For those deployments, no cloud API, including ours, can satisfy the requirement. Self-hosted open-source models or STT vendors that explicitly support on-premise deployment (such as Speechmatics or Vatis Tech) are the correct evaluation path.
  • Legal wants audio controlled, not necessarily on-premise: This is the case for the large majority of enterprise compliance requirements. A team that needs GDPR-compliant processing, HIPAA coverage for PHI, zero data retention after processing, EU-only data residency, and no model training on their audio can satisfy all of those requirements through our cloud API, without managing infrastructure. Zero Data Retention (Enterprise) deletes audio, transcripts, and metadata immediately after callback delivery. EU workloads process and store within EU boundaries by default. DPAs and BAAs are available for GDPR and HIPAA respectively. Customer data is never used for model training on Growth and Enterprise plans. If the underlying compliance goal is control, not literal air-gap, cloud-native governance delivers it, with better accuracy, no hardware overhead, and a vendor SLA behind it.

Routing workloads across models without mixed infrastructure

Teams processing mixed workloads, such as sensitive European contact-center audio alongside high-volume multilingual transcription, don't need a hybrid deployment model to manage them separately. Both workloads run through the cloud API. The routing decision is a model selection decision, not an infrastructure decision.

For European contact-center and business audio in EN, FR, DE, ES, or IT, use Solaria-3: #1 on Earnings22 financial calls, the only model under 7% in that benchmark, and ahead of AssemblyAI, ElevenLabs, Deepgram, Mistral, and Speechmatics on real customer recordings. For high-volume multilingual transcription, code-switching, or real-time streaming, use Solaria-1, which covers 100+ supported languages with custom vocabulary available on all plans to improve recognition of domain-specific terms such as product names, team names, and branded entities. Both models run on the same API with the same data governance controls applied at the account level.

Security requirements for enterprise speech-to-text

Meeting SOC 2 Type II requirements

SOC 2 Type II certification means our security controls are audited continuously over time, not just at a point-in-time snapshot. For speech infrastructure, this covers access controls, audit trails on data processing events, monitoring for unauthorized access to audio files, and incident response procedures.

ISO 27001 adds a framework for information security management at the organizational level, covering risk assessment, asset management, and supplier relationships. HIPAA compliance covers the BAA requirements for healthcare customers processing PHI.

Automated deletion of processed audio

Zero Data Retention (ZDR) is available on Enterprise plans and ensures that audio, transcripts, and metadata are deleted immediately after processing completes, with only billing-essential metadata (request ID, timestamp, processing status, and audio duration) retained in immutable logs for a limited period. The complete per-step mechanism from audio submission through ephemeral processing, callback delivery, and immediate deletion, is covered in the cloud-native data control section below, alongside the encryption and residency controls that sit alongside it. This architecture means there is nothing to breach, subpoena, or accidentally expose after a transcript is delivered.

Securing data in transit and at rest

All data transmitted to and from our API is encrypted in transit via TLS (industry standard). Data stored at rest within our infrastructure uses industry-standard encryption. Access to customer audio and transcript data is isolated per account and continuously audited. Unauthorized cross-account access is prevented at the infrastructure level.

Preventing model training on customer data

This is the clause most operations leads never find until a legal audit surfaces it. Our policy differs by plan, and the distinction matters:

  • Starter plan: Customer data can be used for model training by default on this tier.
  • Growth and Enterprise plans: Customer data is never used for model training. No opt-out action is required and there is no contract clause to locate. This is the default behavior on paid plans.

If your product handles sensitive conversations and you're on a free or Starter tier, you're operating under a different data-use policy than Growth or Enterprise customers. Moving to a paid plan is not just a pricing decision, it's a compliance decision.

Managing global data residency and privacy

EU data sovereignty and GDPR compliance

We're headquartered in Paris with EU infrastructure in France, which means EU data sovereignty is built into our architecture rather than bolted on as an enterprise add-on. EU workloads process and store within EU boundaries by default. Switching to our US cluster requires an explicit configuration change. Data Processing Agreements are available for EU customers under GDPR requirements, covering the lawful basis for processing, data subject rights, deletion schedules, and sub-processor disclosures.

Healthcare teams processing PHI can execute a BAA with us, covering the technical safeguards required under HIPAA. Financial services teams with strict data handling requirements can configure Enterprise with ZDR and per-account isolation, with access continuously audited at the infrastructure level. PII redaction is available as an optional, explicitly configured feature, not enabled by default. When configured, it replaces sensitive entities (names, phone numbers, account numbers) in the transcript with labeled placeholders such as [NAME] or [PHONE_NUMBER]. Do not assume redaction is active unless you have explicitly enabled it in your API configuration.

What cloud-native data control actually looks like

For teams whose compliance requirement is control over audio data rather than literal air-gap, Zero Data Retention on our Enterprise plan achieves the same governance outcome through a verifiably documented mechanism. Here is how the pipeline works end to end:

  1. Audio submission: Customers submit audio via URL (for example, S3 presigned URLs). Under ZDR configuration, direct file upload is disabled. Audio is fetched from your storage and never written to our persistent layer.
  2. Ephemeral processing: Audio is processed in memory without being written to persistent storage at any point during inference.
  3. Callback delivery: The completed transcript is delivered to your specified callback endpoint. No retrievable copy of the transcript or audio is held after delivery.
  4. Immediate deletion: Audio, transcript content, and request metadata are purged on delivery. Only billing-essential metadata (request ID, timestamp, processing status, audio duration) is retained for a limited period in immutable audit logs.

All data in transit uses TLS encryption. Data at rest within our infrastructure uses industry-standard encryption. Access to customer audio and transcript data is isolated per account and continuously audited. Unauthorized cross-account access is prevented at the infrastructure level. Region selection is configured per account.

EU workloads process and store within our EU cluster by default. Switching to our US cluster requires an explicit configuration change. This combination of ZDR, regional pinning, per-account isolation, TLS/at-rest encryption, and continuous audit, is what "data control" looks like in a managed cloud deployment, without any customer-managed hardware.

Defining critical SLAs for production stability

Defining uptime and service credits

We maintain a 99.9%+ uptime SLA across production APIs, with real-time and historical status, including incident history and maintenance windows. A well-documented incident history is a signal of operational maturity, not weakness. Vendors who hide their status pages are hiding their failure modes.

Enterprise SLAs include custom uptime commitments and service credit structures negotiated as part of the annual contract, with service credits applied according to the agreed schedule without requiring manual escalation.

Defining latency budgets for STT

Latency expectations differ significantly between async batch processing and real-time streaming, and the right choice depends on your product architecture.

For meeting assistants and post-call analytics, async batch processing is the correct choice. A 10-minute audio file processes in under a minute, and the full-context analysis that async enables produces meaningfully better accuracy, diarization quality powered by pyannoteAI's Precision-2 (async-only), and multilingual consistency than streaming can provide.

Real-time streaming via Solaria-1 delivers partial transcripts under 103ms and final transcripts around 300ms, which is appropriate for live-assist contact center use cases, voice agents, and live captioning. Solaria-3 is async-only for now.

Defining enterprise response time SLAs

Enterprise customers receive dedicated support through Slack with direct engineering team access, with priority response windows and escalation paths defined at the contract level. Multiple customers independently report resolving integration issues within the same business day through direct engineer contact. Custom Enterprise SLAs include response time commitments scoped per severity level as part of the annual contract.

When self-hosted STT makes sense, and what it actually costs

The narrow case where self-hosted is the only option

Some environments prohibit any external internet connection regardless of encryption or compliance certifications, such as defense contractors, intelligence agencies, and sovereign financial institutions with physical isolation mandates. For those deployments, cloud APIs are structurally incompatible and self-hosted open-source models or vendors with explicit on-prem support are the only viable path. We are not the right vendor for those requirements, and saying so clearly is more useful to you than a sales call that fails at the security review.

Protecting sensitive audio without self-hosting

Executive board meetings, M&A discussions, proprietary trading strategy calls, and privileged legal communications carry risks that go beyond regulatory compliance. For these workloads, the governance requirement is typically zero retention, EU-only processing, no model training on the audio, and a DPA in place, not literal on-premise hardware. Zero Data Retention (Enterprise) means there is nothing to breach, subpoena, or accidentally expose after a transcript is delivered. For teams not operating under a hard air-gap mandate, this is the correct architecture: verifiable deletion, regional pinning, per-account isolation, and continuous audit, with a 99.9%+ uptime SLA and no infrastructure to manage.

Managing infrastructure spend at scale

At very high audio volumes (millions of hours per month), the economics of self-hosting can shift in favor of on-premise even when accounting for hardware and DevOps costs. Teams moving off self-hosted open-source models often discover the actual total cost of ownership is higher than expected: infrastructure overhead, scaling complexity, and the ongoing engineering cost of model upgrades and incident response are routinely underestimated. The Solaria-3 benchmark against competing models on real customer European business audio shows that managed models now outperform self-hosted open-source alternatives on accuracy (#1 on Earnings22) while eliminating maintenance overhead entirely.

How we support complex deployments

Cloud API data control and accuracy verification

Our multi-region cloud deployment supports EU and US regions, with region selection configured per account. Switching clusters requires an explicit configuration change. Zero Data Retention is available on Enterprise and configures audio and transcript deletion immediately after processing. Customer data is isolated per account with access continuously audited.

The blind STT comparison tool strips out provider branding so you pick the better transcript before seeing who produced it. It's a useful gut check, but follow it with a reproducible benchmark on your full audio distribution before making a production commitment.

Security standards and enterprise proof points

Our compliance certifications cover SOC 2 Type II, ISO 27001, HIPAA, and GDPR. DPAs and BAAs are available for regulated industries.

The most direct enterprise proof point for CCaaS deployments is Gravite, a French call quality-monitoring platform that cut call quality review time by 93%, from roughly 15 minutes to 1 minute per call, while transcribing 50,000 hours of audio per year with Gladia. They chose us for French-first accuracy, European data sovereignty, and Solaria-3's benchmark results on telephony audio.

The Solaria-3 launch video details the model's design choices for real-world European business audio, including the trade-offs that make it the right choice for contact-center deployments in EN, FR, DE, ES, and IT, while Solaria-1 remains the correct model for maximum language breadth, code-switching, and real-time streaming.

SLA terms for high-volume deployment

Enterprise SLA terms include: custom uptime commitments above the standard 99.9%+ baseline, dedicated support through Slack with engineering team access, custom model fine-tuning, debundled pricing for organizations with specific per-feature cost requirements, and region selection configured per account for teams with strict data residency requirements. Deployment is on dedicated cloud clusters in EU and US regions. On-premises or air-gapped hosting is not available on any plan.

For teams ready to test the API before committing to Enterprise, start with €50 in free credits and have your integration in production in less than a day. For custom Enterprise deployment discussions including dedicated cluster configuration, ZDR setup, DPA/BAA agreements, and data residency requirements, contact our sales team directly.

FAQs

Does Gladia offer on-premise deployment?

No. Deployment is cloud-only, on dedicated clusters in EU and US regions. On-premises or air-gapped hosting is not available on any plan.

If your underlying requirement is data control, such as zero retention after processing, EU-only residency, no model training on your audio, GDPR DPA, and HIPAA BAA, Zero Data Retention on Enterprise delivers all of those through a documented, audited cloud pipeline without customer-managed infrastructure. Audio is deleted immediately after callback delivery. No retrievable copy exists after processing completes.

Where is audio data processed and stored?

Audio processes in either our EU or US cluster based on your account configuration. With Zero Data Retention enabled on Enterprise plans, audio, transcripts, and metadata delete immediately after processing completes, with only essential billing metadata retained for a limited period.

Does Gladia use customer audio to train its models?

On Growth and Enterprise plans, customer data is never used for model training and no opt-out action is required. On the Starter plan, customer data can be used for model training by default.

What encryption standards does Gladia use?

All data in transit is encrypted using TLS and data at rest uses industry-standard encryption within our SOC 2 Type II audited infrastructure. Access to customer audio is restricted to authorized processing systems under continuous audit.

What is Gladia's documented uptime?

We maintain a 99.9%+ uptime SLA across production APIs, with current and historical status publicly available on our status page, including incident history and maintenance windows.

Is diarization available in real-time deployments?

Speaker diarization powered by pyannoteAI's Precision-2 is optimized for async workflows where full-context analysis produces the highest accuracy. For real-time deployments, speaker attribution can be handled in post-processing.

How long does enterprise integration typically take?

Most product teams integrate the API and reach production in under 24 hours using our lightweight Python or JavaScript SDKs. Enterprise deployment configuration, such as ZDR setup, dedicated cluster region, and DPA/BAA agreements, is scoped individually during onboarding.

Which model should we use for European contact-center audio?

Use Solaria-3 for European business audio (EN, FR, DE, ES, IT), which ranks ahead of AssemblyAI, ElevenLabs, Deepgram, Mistral, and Speechmatics on real production recordings. Use Solaria-1 for maximum language breadth (100+ languages), real-time streaming, and code-switching.

Key terms glossary

Word Error Rate (WER): The standard metric for measuring transcription accuracy. WER is calculated as the number of substitutions, deletions, and insertions required to convert a hypothesis transcript into the reference transcript, divided by the total number of words in the reference. Lower is better. A WER of 10%, for example, means 10 errors per 100 words on a given benchmark dataset and audio condition. Lower figures represent higher transcription accuracy.

Diarization: The process of segmenting an audio recording by speaker identity, answering "who spoke when" rather than just "what was said." Diarization is measured by Diarization Error Rate (DER) and is only available in async (batch) workflows where full-context analysis produces the highest accuracy. Speaker attribution for real-time workflows is handled in post-processing.

Zero Data Retention (ZDR): A cloud-native Enterprise configuration that ensures audio, transcripts, and associated metadata are deleted immediately after processing completes, with no retrievable copy held after callback delivery. Billing-essential metadata (request ID, timestamp, processing status, audio duration) is retained for a limited period in immutable logs. ZDR is how teams with strict data control requirements achieve zero-retention guarantees through a managed cloud deployment, without customer-managed on-premise infrastructure.

SOC 2 Type II: A security certification issued by an independent auditor confirming that a vendor's security controls meet AICPA Trust Service Criteria over a continuous audit period, not just at a point-in-time snapshot. Covers access control, monitoring, incident response, and audit trail requirements relevant to audio and transcript data handling.

Data Processing Agreement (DPA): A legally binding contract required under GDPR between a data controller (your organization) and a data processor (the STT vendor) that defines the lawful basis for processing personal data, data subject rights, deletion schedules, and sub-processor disclosures. Required for any EU-based product processing voice recordings containing personal data.

Business Associate Agreement (BAA): A contract required under HIPAA between a covered entity (a healthcare organization) and a business associate (a vendor that processes Protected Health Information on its behalf). A signed BAA is a prerequisite for any STT vendor processing audio that contains PHI. Without it, using a cloud STT API for healthcare audio creates direct regulatory exposure.

Protected Health Information (PHI): Any individually identifiable health information, including voice recordings of clinical conversations, patient intake calls, or telehealth sessions, that is created, received, maintained, or transmitted by a HIPAA-covered entity or its business associates. STT vendors processing PHI must have a signed BAA in place and must meet the technical safeguards defined under HIPAA's Security Rule.

Contact us

280
Your request has been registered
A problem occurred while submitting the form.

Read more