Changelog

What shipped, when

Every release we think is worth telling you about. Filter by tag or scroll the full timeline.

What shipped, when

2026

Hakim Faseeh v2 + Hakim Mulhim v3#

Now available: Hakim Faseeh v2 + Hakim Mulhim v3

ModelsAPIDocs
  • Renamed the speed tier to Hakim Saree' v1.3 (سريع, *swift*) so every TTS tier reads as part of one Arabic-named family. The API id hakim-fast-v1 is unchanged; the legacy hakim-flash-v1 alias still resolves.
  • Launched Hakim Faseeh v2 (hakim-v2, فصيح, *eloquent*): premium quality with non-verbal tag support across 41 languages. Pick this tier when naturalness matters more than streaming TTFB.
  • Launched Hakim Mulhim v3 (hakim-v3, مُلهِم, *inspiring*): premium quality plus voice generation from a free-form prompt across 41 languages. The full capability set.
  • Shipped the new POST /v1/audio/voices/design endpoint. Conjure a fresh voice from a natural-language description (Identity / Texture / Scenario), iterate, and save the timbre to your library so subsequent speech calls reuse it exactly.
  • Introduced the Hakim non-verbal tag set: [laugh], [breath], [think], [whisper], [ask], [wow], [hmph]. Tags are honoured on Faseeh + Mulhim and silently stripped on Saree' (graceful degradation, signalled via the new x-hakim-warnings response header).
  • Billed per input character in credits: Saree' 1.0, Faseeh 1.5, Mulhim 2.0. Unsupported tags and prompts are stripped before billing, so ignored text is never charged.
  • New /docs guides for Models, Tags, and Voice design. Public OpenAPI spec at api.tryhakim.ai/docs updated end-to-end.
#

Realtime options + a new website

WebAPIDocs
  • Rebuilt tryhakim.ai on our new design system: new home, pricing, enterprise, models and feature pages, refreshed sign-in and sign-up, and a phone menu built for touch, in Arabic and English, light and dark.
  • Realtime TTS sessions accept two opt-in options for text that arrives in chunks: trailing_silence_ms (0 to 500, 0 for gapless joins) and chunk_chain: false for steadier delivery in very long sessions.
  • Realtime STT keeps an idle session open for up to 5 minutes, and its vocabulary keeps applying after long pauses.
  • Voice cloning: two voices can now share a name without an error, and clone previews play reliably.
  • Docs: new sections on custom vocabulary, silence trimming and the realtime TTS session options.
#

Silence trimming for realtime STT

APIDocs
  • New trim_leading_silence option for realtime STT: we hold audio until speech starts and send it from about 300 ms before the first word, so a long silent lead no longer weakens recognition or vocabulary.
  • Works with pcm16 and mulaw input. Off by default, and billing is unchanged.
#

Cleaner realtime audio

ModelsAPIFix
  • No more clicks in streamed audio: resampling is now seamless from chunk to chunk, and stray ticks are removed from pauses.
  • Steadier loudness through long calls: level creep and clipping late in a session are caught and reset, and voices registered from loud reference clips are levelled at registration.
  • Natural endings: utterances finish on their full tail instead of stopping on the last syllable.
  • Faster recovery when a speech host restarts: requests move to a healthy host instead of returning an error.
  • MP3 output now picks its bitrate from the requested sample rate.
#

Steadier long-form speech

ModelsFix
  • Ordinary sentences are no longer split into short pieces before synthesis, so a sentence keeps its natural rhythm.
  • Long passages stay stable: runaway generations are capped and long inputs continue smoothly from one segment to the next.
  • Numbers and currency amounts in mixed Arabic and English text are read correctly.
#

Security hardening

Security
  • Webhook URLs are validated so they can only point at public internet addresses.
  • Browser access to the API is limited to our own sites.
  • Rate limits and quotas are tied to the real client address, so they cannot be sidestepped with forged headers.
  • Patched high-severity advisories in third-party dependencies.
#

Reliability

APIBillingFix
  • If a speech host fails mid-request, the request is retried once on a different host automatically.
  • Realtime TTS and STT sessions connect faster.
  • Fixed occasional 502 responses on long-idle keep-alive connections.
  • Steadier routing: one slow health check no longer takes a host out of rotation, and recovered hosts return sooner.
  • Fixed: the Business plan could not be selected or upgraded to.
#

Custom vocabulary for realtime STT

APIModels
  • Custom vocabulary for realtime STT: pass terms, notes and hints in session.update so names, brands and acronyms come out right.
  • One shared budget of 10,000 characters per session across all vocabulary fields.
#

Cleaner voice clones

Models
  • Voice cloning cleans the reference recording first, removing background noise and harsh sibilance, so clones from phone or room recordings sound cleaner.
  • Clones also use the reference transcript, which improves pronunciation and likeness.
  • Long generations keep a steady pitch from start to finish.
#

Faster realtime STT

APIModels
  • Realtime STT sessions are warm from the first word: no cold-start delay on the first utterance.
  • Lower end-of-utterance latency in realtime transcription.
  • Realtime STT accepts 8 kHz mulaw telephony audio directly.
#

Docs platform + API reference

DocsWebSecurity
  • Shipped a Next + MDX docs platform at /docs with a catalogue-driven sidebar, prev/next, and edit-on-GitHub across Arabic and English.
  • Moved the interactive API reference to api.tryhakim.ai/docs (Scalar, served by the API itself), the docs sidebar and Developers menu now deep-link there in a new tab.
  • Hardened the web surface with X-Frame-Options: DENY, Referrer-Policy: strict-origin-when-cross-origin, and a narrow Permissions-Policy allowlist.
#

Developers hub + quickstart

WebDocs
  • New /developers page with a live TTS/STT snippet carousel (Node, Python, cURL), 5-up SDK grid, and a 4-step quickstart.
  • Copy-paste snippets pin to hakim-fast-v1.3 and hakim-arab-v2 so the code you read matches the model you call.
  • First synthesis call in under five minutes, verified on a fresh machine by the Phase 13 Playwright smoke.
#

Models page + dialect matrix

ModelsAPI
  • Added a dialect matrix covering 15 Arabic dialects on Hakim Arab v2 plus six additional MENA language tiers.
  • Live latency tickers (TTS 113 ms p95 · STT 90 ms p95) backed by the /api/public/latency BFF with a 15 s Redis window.
  • Voice previews now stream in the dialect you click, no round-trip to the playground.
Billing GA#

Billing GA

BillingWeb
  • Five plans live on /pricing with a monthly / yearly toggle and a first-class enterprise intake form.
  • Annual plans bill at a 20% discount; the toggle persists across pricing revisits in the same session.
  • Subscription state is authoritative from Stripe webhooks, no polling, no drift between the hosted checkout and the dashboard.
  • Per-workspace usage meters expose minutes used, quota, and burn-down in both the dashboard and the public /pricing comparison table.
#

KSA + Qatar regions live

SecurityWeb
  • Riyadh region is now the default for Saudi-registered workspaces, traffic never leaves the Kingdom for synthesis or transcription.
  • Doha region serves Qatar-resident workspaces end-to-end: model replicas, object storage, and audit logs all pinned in-country.
  • A new X-Hakim-Region response header surfaces the region that served each request so you can verify residency in your own observability pipeline.
Hakim Arab v2#

Hakim Arab v2, GA with 15 dialects

ModelsAPI
  • Fifteen Arabic dialect varieties are now in-distribution at GA quality, spanning MSA plus the major regional families (Levantine, Egyptian, Gulf, Maghrebi, Iraqi, Sudanese, Yemeni). The full per-dialect MOS scores live in the dialect-coverage post.
  • Average WER on the in-house MENA eval set drops from 10.2% (v1.5) to 6.8% (v2), with the biggest gains on dialect-heavy audio rather than MSA.
  • Diacritics restoration is now a first-class stage instead of a pre/post hook; the model attends to both restored marks and the bare glyphs.
  • Arabic⇄English code-switching no longer drops characters at language boundaries, a direct ask from enterprise design partners.
Hakim Fast v1.3#

Hakim Fast v1.3

ModelsAPI
  • Streaming TTS p95 down to 113 ms end-to-end across Frankfurt, UAE, KSA, and Qatar regions.
  • 24 kHz default output with 16 / 22.05 / 44.1 kHz fallbacks, negotiated via the Accept header.
  • MP3, Opus, FLAC, WAV, plus raw PCM for teams piping into their own mixer.
#

SOC 2 Type I report available

SecurityCompany
  • Completed SOC 2 Type I readiness with a Big-4 auditor; Type II observation window is open and runs through Q3.
  • Launched a public trust center at /compliance with the report, subprocessor list, and standing data-protection posture.
  • Every subprocessor change now ships with a 30-day notice policy baked into the compliance page, no silent substitutions.
Hakim Arab v1.5 (beta)#

Hakim Arab v1.5 (beta)

Models
  • Real-time STT streaming drops p95 first-token latency to ≤200 ms on Frankfurt and UAE, a 40% improvement over v1.
  • Multi-speaker diarization lands in beta for up to 6 speakers on Arabic conversational audio.
  • Added a telephony-audio input mode (8 kHz, narrowband) tuned for GCC call-center recordings.

2025

#

Frankfurt + UAE regions live

SecurityCompany
  • Frankfurt (eu-central-1 equivalent) is now serving EU-resident workspaces end-to-end, independent Postgres, model replicas, and object storage.
  • UAE region brings inference onto Emirates-local infrastructure; first-party partners in the UAE helped us bootstrap the deployment.
  • Workspaces can be pinned to a region at creation; we refuse the request if the requested region isn't yet active rather than silently falling back.
Public launch#

Hakim is out of stealth

CompanyWeb
  • After eighteen months in stealth, Hakim is publicly available. tryhakim.ai now serves self-serve sign-ups for teams anywhere in the world, with the studio, API, and SDKs all open.
  • Free tier is live, every workspace gets enough monthly minutes to ship a real prototype before picking a plan.
  • Our first cohort of design partners, six enterprises across MENA and Europe, move from private alpha to named customers alongside this launch.
  • We wrote our thesis on Arabic-first voice AI up as a founding letter: why we built Hakim.
#

Voice cloning (beta) with consent-gating

ModelsSecurity
  • One-shot voice cloning from a single high-quality sample; 10 seconds is all you need. Cloning is instant and runs the same inference stack as the catalogue voices, no separate training queue.
  • Cloning a third-party voice requires an explicit, signed consent artifact on file, we won't synthesize a voice we can't trace to a paper trail.
  • Every synthesized clone is watermarked at inference; the API exposes a /detect endpoint for downstream verification.
  • Consent can be revoked per-voice; revocation takes effect globally within minutes and retroactively disables previously-minted API keys that targeted the voice.
Hakim Fast v1.2 · SDK 0.1#

Streaming TTS (v1.2) and the first SDKs

ModelsAPI
  • Chunked HTTP streaming lands for TTS: first audio byte at ~180 ms p95, continuously streamed through synthesis end instead of a one-shot blob.
  • Published official SDKs for Node.js and Python (both at 0.1.0). Both are thin wrappers over the REST API with typed request shapes and native AsyncIterable stream support.
  • The transport is plain HTTP chunked transfer, no WebSocket dependency for the common case. Clients behind strict corporate proxies work out of the box.
Hakim Arab v1 (alpha)#

Hakim Arab v1, first Arabic STT cut

Models
  • First production STT model ships to private alpha, the inaugural member of the Hakim Arab family.
  • Covers MSA plus three dialects at this stage (Levantine, Egyptian, Gulf); the full 15-dialect tree lands with v2.
  • Released to a hand-picked cohort of design partners, customer-support, media, and healthcare teams in KSA and the UAE.
Hakim Fast v1.0 (alpha)#

Hakim Fast v1.0, private alpha

ModelsAPI
  • First production TTS model ships to private alpha: Hakim Fast v1.0. The research-preview v0.9 graduates after a full rewrite of the acoustic decoder.
  • Four MSA voices, two male, two female, each trained on native-read studio corpora rather than scraped web audio.
  • Deployed to three design-partner teams under NDA; their feedback drives the v1.1 prosody pass later in the quarter.

2024

Hakim Fast v0.9 (internal)#

Hakim Fast v0.9, internal research preview

Models
  • First working end-to-end Arabic TTS synthesis from the Hakim stack. Internal demo only, no API, no customers.
  • Proof-of-concept dialect conditioning heads for MSA and Gulf. Performance is uneven on the tail, but the decoder shape that eventually ships in v1 is already in place.
  • Released exclusively to the founding team and a handful of advisors for native-speaker listening sessions.
#

Hakim is founded (stealth)

Company
  • Hakim is founded with a four-person team split across research and platform engineering.
  • The company operates in stealth, no website, no public API, no press, while we build an Arabic-first voice stack from scratch.
  • Founding thesis: Arabic-first voice AI is a category, not a locale. Every model, corpus, and evaluation benchmark is built Arabic-first from day one.