Developer Relations · SpeechifyAI Labs
Luke Oliff is a Developer Relations leader based in the UK. For the better part of a decade he has been working with voice technology, developer tooling, and open-source — improving developer experience for well known brands. He has architected open-source strategy, launched developer communities, built tools, and shipped conversational AI voice prototypes years before mainstream APIs were available. As an engineer at heart, he writes and speaks about voice AI, developer experience, and real-time APIs as a developer would, focussing on utility and experience. He has now joined the SpeechifyAI Labs team, where Simba 3.2 sits at #1 on the independent Artificial Analysis TTS leaderboard and ranks as the #1 real-time voice on Voice Arena — the best-sounding real-time voice at its price.
The voice agent platforms developers actually evaluate in 2026, compared on what a minute really costs once the LLM, speech, and telephony are added. Most headline rates are platform fees, not bills. SpeechifyAI is all-in from $0.07/min.
Luke Oliff
The tools for producing narration without hiring voice actors, compared on cost per finished hour rather than per seat. Studio subscriptions cap downloaded minutes; per-character APIs do not. One hour of audio runs from $0.31 to $40 depending on which you pick.
Luke Oliff
What engineers actually use to add voice to web and mobile apps, and the four decisions that matter: the browser's built-in API versus a cloud one, streaming versus batch, where the key lives, and how to keep audio playing on iOS.
Luke Oliff
Which speech platform to pick when your chatbot has to talk. The decision is not voice quality, it is whether you need a TTS API or a full voice agent, plus how the model reads account numbers and order IDs back to a caller.
Luke Oliff
Create a revocable, budget-capped share link for any agent. Share the URL and the recipient opens it and talks — no account, no embed, no setup.
Luke Oliff
GET /v1/agents/voices now returns a multilingual_languages field listing every language an agent may declare in additional_languages, so language pickers render from the live set instead of a hard-coded list that drifts out of date.
Luke Oliff
The voice-agent widget now reports real failure codes on widget.error instead of always-unknown, and startAgent() connect failures arrive on the promise rejection rather than duplicating across onError and the rejection channel.
Luke Oliff
GET /v1/audio/models now returns endpoints and curated_voices per model, so a model picker can show only valid synthesis routes and reject unsuitable voice selections before the request reaches the server.
Luke Oliff
Generate the same script in English, German, Spanish, French, Italian, and Portuguese with the Speechify TTS API. Pick a simba-3.0 voice for each locale, call POST /v1/audio/speech per language, keep the key server-side.
Luke Oliff
New widget embeds load from cdn.speechify.ai - a floating URL that tracks the current release, plus immutable version pins for change-controlled sites - and the Conversation resource's end_reason field now documents flow_budget_exhausted, a runtime backstop for looping calls.
Luke Oliff
The voice-agent widget is now served compressed - 676 kB down to 190 kB when it was measured - supports version-pinned URLs with subresource integrity, and session creation returns 402 when a spend limit or budget is hit rather than a 500.
Luke Oliff
POST /v1/audio/speech, /stream, and /stream/with-timestamps now return voice_not_found or a model-not-found message on a missing voice or model, instead of an opaque passthrough 404.
Luke Oliff
The AI SDK ships generateSpeech() but no Speechify provider. So we wrote one: a dependency-free custom speech model that maps POST /v1/audio/speech onto the SDK's SpeechModelV4 interface, key held server-side in a Next.js route.
Luke Oliff
Free-tier /v1/audio/* requests get a 10-request burst bucket instead of capping at the 1 req/s sustained rate, so a first quickstart script no longer 429s on its second call.
Luke Oliff
Speechify-Request-Id and RateLimit-* are now the canonical header names across the public API; the old X-prefixed spellings keep working until 2027-07-24.
Luke Oliff
One agent serves multiple languages in a single session, with no transfer and no dropped context - configure it with additional_languages.
Luke Oliff
simba-3.2 is now the recommended Speechify TTS model, and GET /v1/audio/models lets you discover it at runtime instead of hardcoding a model list. What shipped, the voice allow-list, and how to move off simba-3.0.
Luke Oliff
The top 10 text-to-speech APIs of 2026, benchmarked on the independent Artificial Analysis Speech Arena. Simba 3.2 is #1 on Artificial Analysis at $6 to $10 per million characters, above every ElevenLabs, Cartesia and Google model. Self-hosted models left out.
Luke Oliff
A new endpoint streams speech marks alongside audio, so captions and text highlighting no longer need the non-streamed API.
Luke Oliff
All four agent test-run endpoints now document 402 for a depleted balance or an exhausted spend limit, the single-run path finally takes the admission gate its siblings already had, and an explicit over-tier model override is refused instead of being honoured unchecked.
Luke Oliff
Simba 3.0 is no longer English-only — it now officially covers English plus German, Spanish (ES/MX), French, Italian, and Brazilian Portuguese. The same release flags the Simba 1.6 models as legacy and confirms Simba 3.0 accepts cloned voices.
Luke Oliff
Voice agents now carry a max_call_duration_seconds field — a hard per-agent wall-clock cap on a single call. When a call reaches it, the agent ends automatically; null keeps the previous behavior of bounding calls only by your plan's ceiling.
Luke Oliff
GET /v1/voices now accepts type, locale, gender, and model query filters, applied before pagination so pages stay full — so you can fetch, say, only English cloned voices that support Simba 3.2 in a single call.
Luke Oliff
The API limits reference now lists per-plan rate limits (requests per second, plus an Agents burst) and concurrency limits for both the TTS and Agents surfaces, and 429 responses carry a docs_url plus distinct rate_limited and concurrency_limited error codes.
Luke Oliff
Clone a voice from a browser upload, synthesize speech with it, and keep your API key server-side. A working Next.js voice cloning app built on the Speechify API, no GPU required.
Luke Oliff
simba-3.0 accepts cloned and personal voice_id values self-serve on POST /v1/audio/speech and /v1/audio/stream, the same way simba-english does; simba-3.2 cloning still requires manual approval.
Luke Oliff
Stream Speechify TTS as raw PCM, keep the API key server-side, and schedule each chunk into the browser's Web Audio API for low-latency playback.
Luke Oliff
Point Deepgram Voice Agent at the open-source tts-shims OpenAI-compatible proxy to speak with a Speechify voice. The shim answers Deepgram's open_ai TTS request and keeps your Speechify key server-side.
Luke Oliff
Use SSML emotion tags, pauses, prosody, emphasis, and pronunciation aliases to shape Speechify TTS output from one API request.
Luke Oliff
Personal cloned voices can now be synthesized on the curated Simba 3.2 model. Given the model's quality bar, each clone is gated on manual Speechify approval of the voice key, while simba-english and simba-multilingual keep serving clones self-serve.
Luke Oliff
A new GET /v1/audio/models endpoint returns the models you can pass as the model parameter, each with a default and recommended flag plus the languages it supports — so a model picker can be built at runtime instead of hardcoding the list.
Luke Oliff
Simba 3.2 is now available on POST /v1/audio/speech and /v1/audio/stream — a streaming-native Simba 3 model with lower TTFB, richer expressivity, and a curated voice allow-list, recommended for new English integrations.
Luke Oliff
Simba 3.2 is #1 on Artificial Analysis, the independent TTS benchmark. On Voice Arena's blind, listener-voted board it's the #1 real-time voice and #1 on price — the model above it isn't real-time, the nearest at its quality costs 7x more. Nothing you can ship beats it.
Luke Oliff
Turn long-form ePub or Markdown into a single narrated chapter MP3 with a runnable Python demo: chunk on sentence boundaries, synthesize each chunk, stitch with ffmpeg.
Luke Oliff
Speechify API keys can now carry a monthly USD spend limit, returning 402 spend_cap_exceeded once a key reaches its budget for the calendar month.
Luke Oliff
Speechify is now a direct provider in speech-sdk. Switch existing calls over with a factory function and an API key, no new client required.
Luke Oliff
Speechify TTS speech and streaming endpoints now accept output_format values like pcm_16000, ulaw_8000, and bitrate-tuned mp3 variants without changing existing callers.
Luke Oliff
Speechify added rate-budget response headers, global JSON body-size handling, clearer 413 and 415 errors, and pre-auth IP throttling for token-mint paths.
Luke Oliff
Clone a voice from a short sample and use it to narrate dynamic video content programmatically via the REST API.
Luke Oliff
Speechify workspace webhooks now support managed endpoints, delivery history, a resource.action event catalog, combined signatures, and per-endpoint API versions.
Luke Oliff
Speechify API keys now have cleaner scope tiers, service-account inheritance, and usage attribution that makes spend easier to trace back to a workload.
Luke Oliff
Baseten published a case study on the SpeechifyAI voice agent stack. This is our view on why model co-location matters for live calls.
Luke Oliff
Service accounts give Speechify API workloads their own identity, scope ceiling, key rotation flow, usage attribution, and short-lived child keys for agent sessions.
Luke Oliff
Use the Speechify API's word-level timestamps to generate accurate WebVTT captions for your synthesized audio.
Luke Oliff
Build and Agents received a small API naming cleanup on June 28, including not_specified voice gender and several Agents field-name migrations.
Luke Oliff
How to stream audio from the Speechify TTS API in Python using the SDK and native requests. Covers chunked streaming to disk and piping audio to a player without waiting for the full payload.
Luke Oliff
The Build API moved GET /v1/voices from a bare array to a voices object with pagination fields, making the response shape safer for SDKs and future growth.
Luke Oliff
Speechify now supports Idempotency-Key on the API calls most likely to be retried, including calls, batches, purchases, spend paths, and service-account key minting.
Luke Oliff
Speechify APIs now support the Speechify-Version header, so HTTP clients can pin a dated contract while SDKs send their build-date version automatically.
Luke Oliff
A practical guide to synthesizing speech in Node.js using the Speechify TTS API. Covers installation, a basic synthesis call, streaming audio to disk, and what to reach for next.
Luke Oliff
The Build docs now spell out that POST /v1/audio/stream returns chunked raw audio, not JSON, with clearer response content types and codec notes.
Luke Oliff
Speechify Agents added live transcript streaming, take-over actions, per-action RBAC, analytics queries, and saved dashboard APIs for monitoring live voice operations.
Luke Oliff
A small, public repo of runnable Speechify recipes. Pick a folder, drop in your API key, run it. TypeScript and Python today, SDK and native REST side by side.
Luke Oliff
Welcome to the new Speechify developer blog. Expect updates from the AI and labs teams, product news, collaborations, and technical guides for the Speechify TTS API.
Luke Oliff
Speechify Voice Agents is now publicly documented, with guides and API reference pages for agents, conversations, tools, knowledge bases, calls, and monitoring.
Luke Oliff