Gladia turns multilingual audio into timed application data

Gladia provides transcription APIs for live audio and existing recordings, with language handling, speaker separation and additional analysis around the resulting text. Its strongest role is as a component inside another product. The developer receives timed conversation data and decides how to display, retain and act on it.

01 / ProductTranscription is structured conversation data

Gladia's pricing and product description distinguish asynchronous transcription from live streaming. Both routes include core recognition features such as language detection, speaker diarization and word timing. An audio file does not need to become one undifferentiated paragraph: those structures can help an application connect a sentence to a speaker and a moment in the recording.

Gladia's 21 September 2026 announcement describes an acquisition by OVH Groupe, the parent of OVHcloud, while stating that regulatory approvals and closing conditions still apply. The announcement says Gladia will retain its own brand, API and direct customer relationship. Treat the transaction as announced, rather than assuming closing or future infrastructure integrations are complete. Existing product documentation and commercial terms remain the basis for this guide.

The code-switching guide covers conversations that move between languages, including within an utterance. It recommends configuring a small set of expected languages and warns against enabling code switching with an empty language list. That is a practical constraint for a multilingual product. Broad language coverage does not mean that leaving every possibility open produces the best result for a known audience.

The diarization guide assigns numbered speaker indexes in order of appearance. Those labels identify clusters of speech, not verified personal identities. Speaker-count settings are hints rather than hard guarantees. A transcript interface should therefore let an authorized reviewer map a speaker number to a person and correct the assignment without pretending that the API has authenticated who spoke.

Additional analysis can make audio easier to navigate, but it inherits the recognition stage's mistakes. A summary of an incorrectly transcribed amount may look more certain than the underlying audio. Preserve the transcript and timestamps alongside any derived view so users can move from an interpretation back to the material that supports it.

02 / AudienceA useful layer for multilingual products

Consider a support platform whose conversations regularly mix two languages, or a research repository indexing interviews from different markets. The relevant need is not merely a translated summary. It is a record that preserves which language was spoken, when it changed and where a human can check a consequential phrase.

Gladia fits teams willing to build those interactions around an API. The application owner controls who uploads recordings, how text becomes searchable and whether annotations leave the system. That is a different responsibility from buying a complete meeting assistant and accepting its default collaboration model.

It is less attractive for someone who only needs a few personal meeting notes without integration work. It also requires a careful evaluation when the audio contains overlapping speakers, rare names or numbers that feed another system. These are reasons to design a useful test set, not grounds for assuming that any one transcription provider is universally accurate or inaccurate.

03 / WorkflowProposed workflow: a bilingual interview archive

Build a pilot around a small set of interviews for which the team has recording and processing permission. Include the ordinary mix of accents, background noise and language switching rather than selecting only studio-quality speech. Prepare a reviewed reference for important names, product terms and statements that the research team might later quote.

For each recording, identify the likely languages and configure that limited set using the code-switching guidance. Keep the chosen configuration with the job record. If a participant introduces another language, investigate that section explicitly instead of treating the initial setting as proof that the whole interview belongs to the expected pair.

Enable speaker separation where it helps navigation. Use the documented diarization output to show speaker groups, then let a reviewer assign names where the recording supports that assignment. Keep the original group identifiers underneath the display labels. That makes a correction traceable and avoids confusing one interview's speaker zero with the same person in another interview.

In a live companion interface, handle draft and completed utterances separately. The transcript-event guide says partial messages can precede a final message for the same utterance and recommends updating by the returned identifier. Replace the draft as it changes. Do not append every version into the permanent archive, which would create duplicated sentences and distort later search or summarization.

Store source timing beside the text and give reviewers a way to jump to the disputed moment. Then ask a colleague to locate a quoted sentence, correct a speaker assignment and remove a record. Those operations reveal whether the archive is maintainable. A transcription pipeline that can ingest content but cannot revise or delete it is unfinished even when every API call succeeds.

Finally, test the intended retention route. If the organization requires zero data retention, the architecture changes: the retention documentation says this is an Enterprise option and results arrive through callbacks rather than later retrieval. The receiving service must accept and persist authorized results reliably. A prototype built around manually revisiting the provider's playground would not exercise that production behavior.

04 / PricingAudio duration and committed usage shape the bill

The current public prices are US$0.61 per audio hour for Starter asynchronous transcription and US$0.75 for real-time. Growth advertises lower starting rates with an upfront commitment, while Enterprise is a custom annual arrangement. Starting prices should not be treated as the rate every project receives regardless of volume or contract.

The billing migration guide describes a prepaid wallet for Free, Starter and Pro accounts, with manual or automatic top-up. It explicitly excludes Growth and custom accounts from that migration. Its introductory grant is one-time rather than a replenishing monthly free allowance. The public site displays that grant in euros while its rate cards use dollar signs; confirm the account's settlement currency rather than silently converting the two.

The billing explanation counts streamed silence and background noise as audio duration. Distinct multichannel content is billed as separate streams, while identical duplicated content is counted once. For a conferencing integration, leaving audio flowing while nobody speaks can therefore affect cost. Model the actual transport behavior and channel layout rather than estimating from spoken words alone.

05 / DistinctionsLanguage configuration is part of product design

Deepgram is a direct comparison for developers choosing a streaming recognition service. Give both systems the same conversation and inspect final text, event timing and recovery. Comparing a clean English demonstration against a noisy bilingual session would say more about the test design than the provider.

AssemblyAI is useful when the product needs transcript enrichment as well as text. Decide which analysis belongs near the transcription service and which the application should own. For example, a research archive may want a stable verbatim layer before introducing changing summary prompts or classification rules.

Gladia's configurable language route is useful because it lets the team incorporate knowledge about the audience. It also creates a responsibility to maintain that configuration as the audience changes. Treat language settings like part of the application schema: version them, test them and avoid applying one customer's assumptions automatically to every other recording.

06 / QuestionsRetention and live updates deserve explicit answers

Does zero retention mean a shorter deletion timer?

No. The retention guide distinguishes a custom minimum retention value of zero, described as deletion within 24 hours, from the separate zero data retention mode. In that latter mode, uploads and later result retrieval are unavailable and callback delivery is required. Default paid audio and transcript retention is three weeks; API metadata has a longer window. Confirm which setting the contract actually enables.

Will a stable-looking speaker label mislead readers?

A numbered speaker group is a model output that may need correction. Do not infer a verified employee, customer or interviewee solely from its index. Display uncertain assignments conservatively and preserve the ability to revise them when a reviewer listens to the recording.

How does the product recover if a callback fails?

A retention policy and a delivery policy meet at this boundary. Before relying on ephemeral processing, test the receiving endpoint, duplicate delivery handling and an interrupted response using the agreed provider behavior. The application should avoid both losing the only delivered result and creating two records from a repeated notification. The exact retry contract should be confirmed during integration.

07 / DecisionChoose the handling of conversation data

Gladia is a relevant candidate when multilingual audio must become structured, reviewable data inside an application. Run a pilot that includes language switches, speaker corrections and the intended retention setting. The strongest evidence for adoption is a complete path from original audio to a useful, correctable record with a predictable audio-duration bill.