Changelog

​

Changed

• Instant voice limit — Free, Starter and Flow workspaces can keep 1 instant (cloned) voice; Creator and Pro keep 10. The limit counts live voices, so deleting one frees a slot. Over the limit, POST /voices returns 403 VOICE_LIMIT_REACHED (was 429)
• Professional voices — available on Creator (2 voices) and Pro (5 voices); other plans get 403 PVC_NOT_AVAILABLE_ON_PLAN, and a full plan gets 403 PVC_LIMIT_REACHED. You can now create more than one professional voice, and the page lists them all
• Marketplace listing — needs a Creator or Pro plan, a submitted KYC tax form and bank details. Submitting works while they are under review; the voice goes live only after KYC is verified and bank details are active. Otherwise submitting returns 403 MARKETPLACE_REQUIREMENTS_NOT_MET
• Downgrades — existing voices are kept; creating and listing new ones is blocked while over the limit

​

Added

• Get Voice Limits (GET /voices/quota), List Professional Voices (GET /voices/professional/jobs) and Get Marketplace Eligibility (GET /voices/professional/marketplace-eligibility)
• MCP — read-only tools sixtydb_get_voice_quota, sixtydb_list_professional_voices and sixtydb_get_marketplace_eligibility

​

Fixed

• STT auto-detect — omitting languages, or sending null, [], "auto" or ["auto"], now expands to the full language list the STT server advertises. Before, the session started but real speech failed with No compatible route exists for this candidate set. session_started.language_mode is candidate_auto for two or more languages and selected for one
• TTS upstream — the /ws/tts proxy connects to the TTS server through its hostname instead of a fixed IP address

​

Changed

• TTS context_created now includes session_id and binary_audio (always false; audio is base64 in audio_chunk)
• TTS warning and context_aborted are documented. warning with code synthesis_interrupted means new text arrived mid-synthesis; send flush_context to speak it
• Realtime TTS runs without audio enhancement — the proxy sets audio_enhance to false on every create_context, whatever the client sends

​

Fixed

• Documented response fields — POST /tts-synthesize (without stream) now returns success, message, audio_base64, sample_rate, duration_seconds, encoding and output_format, as documented. Previously it returned the engine’s raw audioContent (headerless PCM) and an internal conditioning object
• Playable audio — for LINEAR16, audio_base64 is a complete WAV file (16-bit mono PCM with header), so it can be saved or played directly
• Backward compatible — audioContent is still returned unchanged for existing integrations. Streaming ("stream": true) is unchanged: NDJSON lines with result.audioContent
• No charge without audio — a response that carries no audio now returns an error and is not billed. See Text to Speech

​

Changed

• Audio projects — create a project, write your text and generate narration. Import and URL import were removed
• Smart regeneration — the first generation converts every line; after that only new or edited lines are generated (and charged). Lines that need generating are marked, and “Generate all” skips finished blocks and respects your plan’s concurrency limit
• Download — the whole project downloads as one MP3 (GET /studio/projects/:projectId/download). Missing lines are generated first; the endpoint returns 409 AUDIO_INCOMPLETE with the missing lines otherwise
• Emotion tags — insert [happy], [laugh], [pause 1s] and the rest from the Emotion picker, or use Enhance / Remove tags, the same as Text to Speech
• Billing — Studio generation is always charged to the workspace you are signed in to, at the TTS rate

​

Changed

• Upload only — the professional clone wizard is now Intro → Upload → Verify → Review. The teleprompter recording method was removed
• Marketplace listing is optional — choose it on the Review step (off by default). Unlisted voices are ready to use privately right after training, with no review. POST /voices/professional/jobs/:id/finalize accepts list_in_marketplace (boolean); jobs return list_in_marketplace and the new status private
• Submit later — POST /voices/professional/jobs/:id/submit-to-marketplace sends a private voice to marketplace review (private → awaiting_approval) without retraining. See Professional Voice Clones

​

Changed

• Concurrency limits now depend on your plan and apply per workspace (REST + WebSocket combined, shared by members and API keys). TTS: Free 1, Starter 5, Creator 10, Pro 20. STT: Free 2, Starter 10, Creator 20, Pro 50. Previously a fixed per-user cap (TTS 5, STT 8)
• 429 / 1008 errors — TTS_CONCURRENCY_LIMIT / STT_CONCURRENCY_LIMIT now name your plan and its limit; details adds plan
• API — GET /plans adds concurrency { tts, stt } per plan; GET /billing/current-plan adds concurrency for the workspace. See Concurrency limits

​

Added

​

Annual billing

• Yearly prices — Starter $60/yr ($5/mo), Creator $220/yr ($18.33/mo), Pro $990/yr ($82.50/mo). Yearly is monthly × 10, so you get 2 months free. Credits stay a monthly allotment (120K / 400K / 2M per month) and reset every month on your subscription start day
• Subscribe on a cycle — POST /billing/subscribe takes billing_cycle: "monthly" | "yearly" (default monthly) and returns action (subscribe, upgrade or switch_to_yearly), billing_cycle and amount. See Subscribe to a Plan
• Plan prices — GET /plans adds pricing.yearly, yearly_monthly_equivalent and yearly_savings_pct. See List Plans
• Current plan — GET /billing/current-plan adds data.plan.billing_cycle, next_credit_reset_at and next_payment_at. For yearly plans next_payment_usd is the yearly price. See Current Plan
• Plan changes — monthly to yearly and upgrades take effect immediately, and unused plan credits carry over. Yearly to a higher yearly plan gets a prorated discount at checkout. Downgrades and cancellations take effect at the end of the billing cycle
• Errors — 403 YEARLY_TO_MONTHLY_NOT_ALLOWED (cancel the yearly plan, then choose monthly), 400 DOWNGRADE_KEEPS_CYCLE, 400 CYCLE_NOT_AVAILABLE, 400 INVALID_BILLING_CYCLE

​

Updated

• Plans & credits guide — new Annual billing section with the price table, plan-change rules and refund policy. Top-ups on yearly plans use the monthly plan rate

​

Added

​

Pay-as-you-go top-ups (Creator and Pro)

• Buy top-ups — Creator and Pro workspaces can buy $10–$1,000 of credits (presets $10, $25, $50, $100) with POST /billing/add-funds (purpose: "credits"). Top-up credits never expire and stay usable after a downgrade
• Plan rate — credits = floor(usd × plan monthly credits / plan price). Creator buys about 18,181.8 credits per $1 ($10 → 181,818), Pro about 20,202 per $1 ($10 → 202,020). The credits are quoted at checkout and fixed, even if you change plan before the payment completes. Coupons lower the price, not the credits
• Quote and balance — GET /billing/topup returns eligibility, your plan rate, preset quotes and the plan / top-up credit balance. GET /billing/topup/quote?amount= previews a custom amount
• Dashboard — Subscription → Platform has a new Top-up credits card with a Pay-as-you-go button. When credits run out, the You’re out of credits dialog offers Pay-as-you-go (Creator, Pro) or Upgrade to Creator (Free, Starter) instead of an error toast
• Ledger — GET /billing/usage-logs lists top-ups as TOP_UP rows, and SA_TOP_UP for offline payments recorded by 60db support. AUTO_TOP_UP is reserved for automatic top-up
• Errors — 403 TOPUP_NOT_AVAILABLE_ON_PLAN on Free and Starter, and 400 TOPUP_AMOUNT_OUT_OF_RANGE outside $10–$1,000

​

Updated

• Out-of-credits 402 — the message now points to a top-up (Creator, Pro) or an upgrade (Free, Starter), and the body includes details.topup_available and details.auto_topup_enabled
• Plans & credits guide — new Plans, Credits & Top-ups page. Usage draws plan credits first, then top-up credits
• Pricing pages — Memory and Judge pricing no longer describe a pay-as-you-go wallet only

​

Coming soon

• Automatic top-up — save a card once and credits reload automatically when your balance falls below a threshold you choose, up to a monthly limit you set. Not available yet

​

Added

• Emotion tags — put [happy], [calm], [apologetic] and 17 other emotion tags inline in text to change how the following sentences sound, blend two with [sad, calm], and add sound tags such as [laugh], [sigh], [gasp] and [pause 500ms]. Unknown tags are dropped, never spoken. Works on POST /tts-synthesize and in the dashboard. See Emotion Tags
• Enhance — POST /tts/enhance adds emotion and sound tags to your text (up to 5,000 characters) and changes nothing else: your words, spacing and language come back exactly as sent. No credits are charged. See Enhance Text
• Dashboard — the Text to Speech page highlights tags as colored chips and adds Add emotion, Enhance, Undo enhance and Remove tags
• Audio enhancement — generated audio is now polished by default. Send audio_enhance: false on POST /tts-synthesize to skip it (the older enhance still works). It is never applied to the real-time WebSocket TTS used by voice agents, so live calls keep their low latency. The dashboard has a new Audio enhancement switch. See Text to Speech

​

Added

​

Click-to-Call (Production)

• Place a call — POST /dialer/click2call rings the agent on your own DID, then the customer, then bridges them. No SIP client, no WebRTC, no microphone. Returns 201 when a call was dialled and 200 when the reference_id was already bound to a live call and nothing new was dialled (data.placed / data.replayed)
• Read a call — GET /dialer/click2call/{id} by call_id or your own reference_id; ?refresh=1 polls the platform for the authoritative state and repairs a call whose terminal webhook never arrived
• History — GET /dialer/click2call with limit/offset, mine=0 for the whole workspace, and api_key_id for a per-integration statement
• Signed webhooks — pass callback_url + callback_secret (16–256 chars, yours to choose) per call; every verified event is relayed to you with x-qcall-signature: v1=<hmac-sha256> over timestamp + "." + raw_body. Up to 6 attempts, ≥2 min apart, abandoned after 24 h
• Delivery inspection — GET /dialer/click2call/{id}/events shows every event, whether it was applied, and the status your own endpoint returned for it
• Recordings — record: true on placement, then POST /dialer/click2call/{id}/recording for a short-lived playback URL (audio uploads after the call ends — check recording_available)
• Readiness — GET /dialer/click2call/config reports whether the server can dial, whether it can verify results, and the numeric limits it enforces
• Idempotency — reference_id is optional and minted per call when omitted; it is permanent per trunk, and a failed call still consumes it. Refusals that placed no call (429, 503) are safe to retry with the same one
• Billing — ₹0.30/min ($0.003/min) on bridged talk time only, prorated per second. Unanswered calls are free. cost_usd and billing_status are on the call itself
• Limits — 20 calls/min per caller (C2C_RATE_LIMITED), ring timeouts 5–60 s (defaults 30 agent / 45 customer), max_duration 30–7200 s (default 3600), metadata ≤ 2 KiB
• See the Click-to-Call guide and API reference

​

Updated

• Dialer feature guide — now distinguishes the two ways to place a call: POST /dialer/calls (your SIP client dials) vs POST /dialer/click2call (60db dials both legs)
• Introduction — Click-to-Call feature card

​

Added

​

Music API (Beta, Free)

• Song generation — POST /songs (simple mode: one prompt, we write lyrics; advanced: full control)
• Song polling — GET /songs/{id} (async generation, ~1.5–2.5 min)
• Voice management — catalog voices or save custom voices from audio samples
• Download — GET /songs/{id}/download (MP3)
• SDK methods — client.music.create(), list(), get(), update(), delete(), download(), waitForCompletion(), listVoices(), createVoice()
• MCP tools — sixtydb_music_* tools for Claude Desktop
• See Music feature guide and API reference

​

Dialer API (Production)

• Number management — rent Indian numbers (E.164, +919876543210), set caller ID, release
• Call reservation — POST /dialer/calls (SIP client places the actual call; API reserves)
• Call history — GET /dialer/logs (CDR with billsec, disposition)
• Recordings & transcripts — auto-record, fetch playback URL (~5 min), get transcript
• Webhooks — subscribe to call/recording events (create, test, rotate signing key, delivery history)
• Usage/billing — track charges per-minute and monthly subscriptions
• Setup prerequisites — trunk + KYC done in dashboard (not API; gates are 409/403 on missing setup)
• Pricing — ₹600/month ($6) per number; ₹0.30/min ($0.003) per call (inbound/outbound, prorated/sec)
• SDK methods — client.dialer.status(), numbers.search()/buy()/list()/release()/setCallerId(), calls.list()/get()/reserve(), recordings.list()/playbackUrl()/transcript()/delete()/setEnabled(), billing.subscriptions()/usage(), webhooks.list()/create()/update()/delete()/rotateSecret()/test()/deliveries()
• MCP tools — sixtydb_dialer_* tools; confirm: true required for operations that charge money or release resources
• See Dialer feature guide and API reference

​

Updated

• SDKs — Music and Dialer sections added to JavaScript and Python docs
• Introduction — feature cards for Music and Dialer
• MCP Server — new music and dialer tool groups

​

Changed

• speech_final now always equals is_final. Every committed result is a single transcription with is_final: true (and speech_final: true); interims carry both false. There is no second “canonical” emit. Simple consumers append each final and clear the interim.
• config.audio_enhancement values are now "on" | "off" (default off). The old "light" / "adaptive" (and boolean true) are accepted for compatibility and map to "on". Use "on" for noisy / telephony audio.
• Billing now charges every committed final by its duration. Withheld finals (empty text) and no_speech are never charged. A single turn can produce several finals (short-pause splits, segment_final on long monologues) — their durations are non-overlapping, so there is no double-charge.

​

Added

• Turn-shaping config: early_final (bool, default true; send false to keep whole sentences in one final) and turn_detector (bool, default false; send true to wait until the caller finished the thought — emits turn_complete). See the three latency/accuracy profiles in the STT reference.
• More config: interim_min_voiced_ms (default 600), interim_delta_streaming, token_streaming, max_utterance_sec (3–30, default 30).
• finalize client message ({ "type": "finalize" }) — closes the current turn now and returns its transcription (or no_speech); the session continues.
• New server events: no_speech (turn had no words; terminal true/false), turn_complete (only with turn_detector: true), interim_delta (with interim_delta_streaming), transcription_tokens (with token_streaming), and warning (e.g. audio_format_assumed).
• New transcription fields: segment_final (length-cap split — check before speech_final), final_id (unique per committed result — de-duplicate on it), and endpoint_reason (vad_silence | client_finalize | client_stop | max_utterance | language_change | response_timeout | interim).
• Language coverage is now 45 (adds ja, ko, zh, vi, tr, he, and more). GET /v1/languages is authoritative.

​

Removed

• config.continuous_mode — streaming is always continuous; the field is ignored (a session no longer stops after the first utterance).
• config.remove_fillers and config.no_speech_threshold — no longer used (filler stripping was an LLM-refinement feature, which is gone).
• The two-phase / refined protocol and its fields: the separate refined event, is_final: true, speech_final: false first emits, llm_applied / llm_reason / llm_latency_ms, tentative / tentative_reason, and session_started.llm_refinement.
• The mode_changed event (it acknowledged continuous_mode).
• The old skip-mode markers (speech_end_no_result, speech_end_too_short, hallucination_rejected) — withheld low-confidence finals now arrive as text: "" with processing_mode: "low_confidence_transcript".

​

Migration

• Drop continuous_mode / remove_fillers / no_speech_threshold from start.config; add early_final: false and turn_detector: true for whole-sentence finals (profile A).
• Change audio_enhancement: "adaptive" / "light" → "on".
• Gate on any is_final: true (equivalently speech_final: true); stop waiting for a second canonical emit and remove refined-event handling.
• words[] timings use probability, not confidence.

​

Changed

• The upstream STT server rejects a start carrying a lower value instead of clamping it, which killed the session outright.
• The 60db /ws/stt proxy now clamps any value below 1000 ms up to 1000 ms before forwarding — on start and on mid-session config — so existing clients sending 500 keep working and simply get 1000 ms behaviour.
• Need faster turn-taking? Drive barge-in from speech_started and interim results rather than lowering this value.

​

Fixed (docs)

​

Documented

• session_started.llm_refinement — whether your context actually opened the refinement gate.
• llm_applied, llm_reason, llm_latency_ms on canonical transcription emits.
• tentative / tentative_reason — proxy-substituted text when upstream rejects an utterance as a hallucination.
• The proxy’s word-preservation guardrail: a refinement that keeps under 40% of the first emit’s words is rolled back to the original text.
• connection_closed, the AUTH_FAILED error code, and the nested mid-session INSUFFICIENT_CREDITS error shape.
• Billing: only canonical finals are charged; speech_end_no_result, speech_end_too_short, hallucination_rejected and low_snr_dropped never are.

​

Changed

• audio_config: { audio_encoding, sample_rate_hertz } — nested block matching the upstream schema.
• timestamp_type (default "WORD") — enables per-word timestamps in the final NDJSON chunk.
• Legacy flat keys (sample_rate, audio_encoding) still work; new clients should use audio_config.

​

Added

• New boolean field inside start.config. Default false.
• When true and context is set (LLM refinement gate is open), the canonical transcription emit (speech_final: true) drops filler words — um, uh, like, you know, …
• No effect when context is omitted (refinement stays off).
• The 60db /ws/stt proxy coerces non-boolean values to false rather than rejecting the session.

​

Added

• Now accepts default_voice_id (string), marketing_consent (boolean), consent_timestamp, and consent_source — all optional and backward compatible.
• Marketing consent is server-stamped: consent_timestamp / consent_source are set on the server; any client-sent values are ignored.
• Consent is persisted best-effort after the onboarding commit, so a consent-write failure can never fail the onboarding save.

• Public, HMAC-token-verified endpoint for lifecycle email. Sends a List-Unsubscribe (one-click) header for bulk-sender compliance.

​

Changed

• Transactional and lifecycle mail send via Cloudflare (smtp.mx.cloudflare.net, implicit TLS on 465) authenticated with a Cloudflare API token; DKIM/ARC are applied automatically. Configure CLOUDFLARE_API_TOKEN (used as the SMTP password) and SMTP_EMAIL (a sender on a domain onboarded for Email Sending).
• Signup verification email is now sent after the account is committed, so a transient email/transport failure can no longer roll back account creation.

​

Fixed

• A workspace with more than one active membership row no longer appears multiple times in the response; the list is de-duplicated per workspace.

• Submitting onboarding no longer returns 500 when the marketing-consent columns are absent from the database (consent persistence is best-effort — apply the consent migration to enable it).

• SMTP / Cloudflare credentials are scrubbed from any thrown or logged email error.

​

Security

• Cerbos authorization checks added across additional routes (STT, TTS, SLM, Studio, webhooks, sixtydb).

​

Migrations

• New environments must apply the consent columns on users (marketing_consent, consent_timestamp, consent_source); plus the events / lifecycle-email tables if the lifecycle email program is enabled.

​

Added

• New STT transcribe API powering the STT editor.
• Audio playback in the STT editor.

​

Fixed

• Resolved an audio-URL leak.
• Fixed audio not playing after a redirect.

​

Added

• Coupon system for wallet top-ups — percentage / fixed discounts and deposit multipliers (e.g. 2× / 3×). Redemptions are recorded for audit (coupon_redemptions, referencing the top-up’s payment_transaction_id).

​

Added

• Server-side request timeout raised from 120 s → 600 s; clients should match. The 25 MB / ~1-hour file-size cap remains.

• New keywords form field on POST /stt. CSV with optional per-term weights, e.g. Acme:5,XYZ Pharma:8.
• The same boost runs on the WebSocket path via the existing context.terms field — no client change required to benefit; existing terms payloads now produce boosted / original markers automatically.
• Words replaced by the boost appear in the response with boosted: true and original: "<pre-boost word>".

• return_timestamps=word adds per-word start / end to the REST response.
• include_confidence=true adds per-word confidence (0-1).

• min_speakers / max_speakers form fields on POST /stt.

• script_correction=true for code-mixed Devanagari / Latin audio.

• snr_db is now a top-level field on every STT response and a per-message field on WS transcription events. Surface as a “good / fair / poor” badge (>= 15 good, 0–15 fair, < 0 poor).

​

Limits

• REST — 429 with error_code: STT_CONCURRENCY_LIMIT / TTS_CONCURRENCY_LIMIT and details.limit.
• WebSocket — error frame followed by close code 1008. The cap releases when an in-flight request completes; no server-side queueing.

​

New error codes

​

New warning codes

​

Billing changes

• WS double-bill fix. Sessions that previously double-billed when the client sent stop are now billed correctly. billing_summary.total_duration_seconds will show as ~50 % of historical values for affected sessions.
• billing_summary.client_estimated_seconds (new, diagnostic) — rough estimate of audio the client sent. Useful only for debugging duration drift; never display as a billed number.
• Low-SNR refunds. REST sessions where upstream dropped audio for low SNR no longer charge the user. WS already handled this.
• Cancellation propagation. Client disconnects mid-request now correctly abort the upstream call. No charge for cancelled requests (499 STT_CLIENT_CANCELLED).
• Diarization surcharge documented. When diarize=true, requests incur a documented +30 % surcharge on top of the base STT rate to cover GPU diarization (pyannote).

​

Plan-tier gate

​

Surface summary

• REST — POST /stt accepts new form fields (keywords, languages, min_speakers, max_speakers, return_timestamps, include_confidence, script_correction, min_split_sec); returns new fields (snr_db, language_source: "long_audio_chunked", per-word boosted/original/confidence, per-segment chunk_idx).
• REST POST /tts, /tts-synthesize, /tts-stream — may now return 429 TTS_CONCURRENCY_LIMIT.
• WebSocket /ws/stt — adds low_snr_dropped to processing_mode; per-word boosted/original; message-level snr_db; concurrency-limit error frame + close 1008; corrected billing_summary with new client_estimated_seconds field.
• WebSocket /ws/tts — concurrency-limit error frame (legacy {error: {message, code, details}} shape) + close 1008.
• Web UI — axios timeout raised to 600 s, error-code-aware toasts, warning_codes surfaced, WS 1008 distinguished from auth/network failures. The 25 MB upload cap is unchanged.

​

Added — Context-gated LLM refinement

​

REST — context: string

curl -X POST https://api.60db.ai/stt \
  -H "Authorization: Bearer $API_KEY" \
  -F "[email protected]" \
  -F "context=Cricket coaching session. Players: Arjun Mehta, Ishaan Verma. Discussing batting technique, stamina, running between wickets."

await client.speechToText(file, {
  language: "auto",
  diarize: true,
  context: "Cricket coaching session. Players: Arjun Mehta, Ishaan Verma.",
});

client.speech_to_text(
    audio_file,
    language='auto',
    diarize=True,
    context='Cricket coaching session. Players: Arjun Mehta, Ishaan Verma.',
)

60db stt:transcribe --file meeting.wav \
  --context "Cricket coaching session. Players: Arjun Mehta, Ishaan Verma."

{
  "audio_url": "https://example.com/meeting.wav",
  "context": "Cricket coaching session. Players: Arjun Mehta, Ishaan Verma."
}

​

WebSocket — structured {general, text, terms}

{
  "type": "start",
  "languages": ["en", "hi"],
  "context": {
    "general": [
      { "key": "domain", "value": "Cricket coaching" },
      {
        "key": "players",
        "value": "Arjun Mehta, Ishaan Verma, Aryan Khan, Rohan"
      }
    ],
    "text": "Coach reviewing a batting practice session.",
    "terms": ["Arjun Mehta", "Ishaan Verma", "off-side", "stamina", "wickets"]
  },
  "config": {
    "encoding": "linear",
    "sample_rate": 48000,
    "continuous_mode": true
  }
}

​

Changed — WebSocket two-phase canonical flow

1. First emit — is_final: true, speech_final: false — fast dict-corrected text. Use for low-latency UI paint and voicebot barge-in / NLU.
2. Canonical — is_final: true, speech_final: true — definitive answer. Either LLM-refined (when llm_applied: true) or the original text re-emitted (when LLM was skipped or failed).

​

Added — Word-preservation guardrails (proxy + UI)

• Refined word-retention guardrail. If the canonical text drops more than ~60% of the first-emit’s token count, the proxy rolls back to the first-emit text and marks llm_applied: false, llm_reason: "dropped_too_many_words". Tuned so legitimate polish (filler removal, 10–40% compression) passes through untouched.
• Hallucination-rejected fallback. When the upstream server’s word-rate guard emits an empty final (processing_mode: "hallucination_rejected", common on fast-paced Indic/English mixed audio), the proxy upgrades it into a tentative canonical using the cached first-emit / last-interim text plus tentative: true and tentative_reason: "hallucination_rejected" so words never silently disappear.

​

Added — Backward-compat refined event handling

​

Surface summary

• REST — POST /stt, POST /v1/transcribe — context: string form field.
• WebSocket — /v1/stream, /ws/stt — context: {general, text, terms} on start; two-phase canonical flow; llm_applied / llm_latency_ms / llm_reason on canonical.
• JavaScript SDK — client.speechToText(audio, { context }).
• Python SDK — client.speech_to_text(audio_file, context=...).
• CLI — 60db stt:transcribe --context "...".
• MCP Server — sixtydb_stt_transcribe accepts context.
• Proxy — word-retention guardrail, hallucination fallback, first-emit caching, legacy refined shim.
• Web UI — context input in Speech-to-Text page and realtime demo; segments rendered with tentative marker when flagged.

​

Environment knobs (server-side)