Multilingual speech recognition can fail not because a language is unsupported but because the application silently locks onto one language. Gladia documents explicit code-switching controls and a different delivery model for recorded files and live audio.
A bilingual utterance is not two separate files
Gladia’s October 2025 support guide documents a language_config object with a languages list and a code_switching boolean. When the spoken language may change during the same recording, code_switching should be enabled deliberately. Restricting the candidate languages to those actually spoken can reduce misclassification. A generic claim that a product supports more than one language says little about mixed-language sentences, proper names or dialect switching mid-word.
Recorded and live pipelines differ
The vendor publishes distinct asynchronous transcription and real-time WebSocket paths. The recorded path is well suited for longer editorial processing and enrichment; the live path supports intermediate events that may be revised as speech arrives. The 2025 WebSocket guide describes creating a session and binding a single socket to it. Do not copy older V1 examples into a V2 integration without verifying the JSON shape and endpoint. Store the API version in each job.
Treat partial transcripts as provisional
Gladia’s support guide explains that live transcript events include an is_final flag. Applications should update text for repeated partial events until the utterance is final, rather than appending every partial as a new sentence. A subtitle preview can show provisional text, but durable customer records should wait for final output and an audit step. Silence endpointing and latency behavior need testing in the actual client environment rather than relying on a vendor demo.
Diarization and confidence caveats
Who-spoke-when and which-language-was-spoken are separate tasks. Gladia documentation covers speaker diarization and timestamps in specific workflows, but developers should verify which features work in the chosen asynchronous or live API version. Mixed accents, overlapping speakers and noise can cause inconsistent boundaries. Report true-language transcription accuracy separately from language identification and speaker segmentation, and retain an ambiguous label where warranted.
A realistic multilingual workflow
A customer-interview team records English and French in one session. Configure known languages, enable code switching, save a tokenless session ID in the app logs and collect final transcriptions. An authorized bilingual editor audits technical terms, names and language change points. Generate captions as a second artifact with reading-speed constraints rather than trusting raw word timing. Respect data-classification rules before sending private interviews to an external service.
Pricing, permissions and date scope
All changing statements are scoped to documentation consulted by 25 October 2026. Model and account support, rights, language coverage, endpoint configuration and regional access can change. This article does not assert first-hand testing, an independent benchmark, present-day tariff, guaranteed authorship or unconditional commercial rights. Recheck the exact subscription, license and supported region before production.