Changelog

Previous

API-breaking changes, new fields, deprecations.

2026-11-09

Realtime refinement is available with Lite and Pro profiles

  • session.update.refinement opts a session into per-utterance AI refinement. true selects the default lite profile; "lite" / "pro" select a profile explicitly; false disables it. It is off by default and fixed before audio; later changes return session_features_locked.
  • Lite uses a one-sentence prompt, two segments of context, and tight character/completion budgets. It is included in the base transcription price.
  • Pro uses the full prompt, ten segments of context, and a rolling guide. It is billed at +$0.012/min on top of the realtime base rate.
  • session.created.features.asr_correction reports whether refinement is available on the gateway; features.refinement reports the enabled profile id (false, "lite", or "pro").
  • New gateway.correction.configured lifecycle event records the profile and prompt version.
  • Rate card published. GET /v1/pricing now returns rate_card_status: "published"; the pricing preview banner is removed.
  • See Pricing for the current rates.

2026-11-08

Recorded translation is available

  • target_languages (array, max 8) and translate_to (single-language shorthand) now work on POST /v1/audio/transcriptions/jobs. The target set is frozen at job creation.
  • New result.translations[] — one track per target language, with per-segment id, timestamps, translated text, and source_text.
  • New result.metadata.translation — model, failed_batches, and token usage (calls, prompt_tokens, completion_tokens, total_tokens).
  • New result.metadata.llm_usage — whole-job token usage across refinement and translation, including a by_stage breakdown. Realtime usage events also record llm_usage per flush.
  • Pricing. Recorded translation is billed per target language × processed audio duration. See Pricing.

Documentation naming and contract clarification

  • Lanson Live API is the developer-facing speech capability; Lanson Live App is the end-user product. LansonAI remains the company and platform brand.
  • Lanson Live is the speech model brand, with model name lanson-live-1. Integrations do not need to select a model; the optional HTTP compatibility parameter can be omitted. Existing client request values and API paths are unchanged.
  • Trinity Engine, StableStream™, and LiveFinal™ now have distinct engine, output-technology, and correction/finalization roles. New LiveFinal™ documentation maps the mechanism to current public events.
  • The public Realtime endpoint delivers one completed result per successfully processed utterance. Optional refinement happens before delivery; no partial/revision or session-wide final event is implied by the technology names. Manual flush is an accepted compatibility message, not an implemented finalization command on this endpoint.
  • These are documentation corrections, not a newly released model-routing or event protocol.

2026-11-07

Realtime translation is available

  • session.update.target_languages (and the translate_to shorthand) enables realtime translation. Up to 8 BCP-47 targets per session.
  • Session-fixed features. The first accepted target set is frozen before audio; later changes return session_features_locked. Start a new session to change targets.
  • New session.updated event confirms the applied language, target_languages, and features.
  • New live.translation.completed event carries one translation per utterance and target language. It arrives independently of the source transcript.
  • Pricing. Realtime translation is billed per target language × processed audio duration. Refinement is now also separately billed when enabled. See Pricing.
  • Recorded translation was added on 2026-11-08 (see above).

2026-11-06

Documentation corrections

  • Realtime language is optional. All examples now show the recommended explicit code, while documenting that omitting it (or auto) enables auto-detection. Passing a known language is recommended for accuracy.
  • Mid-session language switching. A session.update with "auto" resets detection and re-detects on subsequent speech; a concrete code always wins. This supports sessions that start in one language and switch after another appears.
  • Two auto modes documented. Clients can detect first with POST /v1/audio/language-detection and send an explicit code, or omit the language and let the session use its internal detection strategy.
  • Additive event fields. conversation.item.input_audio_transcription.completed now exposes language_source, language_provisional, and optional language_detection_error (also mirrored inside transcript). Existing fields are unchanged.
  • Language detection is documented. New POST /v1/audio/language-detection reference for clients that do not know the language up front.
  • Binary frames are not supported today. The realtime examples and the server-side guide now use base64 JSON input_audio_buffer.append frames.
  • Segment language is required. Missing or auto returns 400 invalid_language.
  • Batch review defaults to on. Documentation now matches the OpenAPI schema: omitted = on, review: false opts out.
  • Server event examples updated for the current completed payload (raw_text, stt_endpoint, worker_id, transcript) and session.created.features.
  • Corrected confidence/timestamp examples (confidence: 0 when unavailable, timing: "estimated" on the batch path) and clarified that SRT/VTT generation is client-side.

2026-08-22

API paths

  • Batch jobs — preferred enqueue: POST /v1/audio/transcriptions/jobs (JSON audio_url, 202).
  • Poll — preferred: GET /v1/audio/transcriptions/jobs/{workflow_id} (opaque id: local + 32 hex on the production endpoint, cf_ + 64 hex on the Cloudflare Worker deployment).
  • Segment — POST /v1/audio/transcriptions is now documented as multipart-only synchronous client-VAD transcription (200).
  • Deprecation — POST /v1/audio/transcriptions with JSON still works but returns Deprecation: true and Link: </v1/audio/transcriptions/jobs>.

Documentation