Changelog

Newest first. Development environment.

22 November 2026

Agent 0.47.0: SignalWire and Telnyx carriers, premium voice by default

  • Your pay-once agent can now use SignalWire or Telnyx as its phone and text carrier, alongside Twilio. The Connect tab has a card for each, with a carrier picker and a step-by-step guide with screenshots.
  • Every carrier now speaks in your agent's premium voice by default. If the premium voice is missed on a call, the agent falls back to the carrier voice and logs it.
  • Calls that are not live now play at the speech rate you set.
  • For carriers other than Twilio, the carrier status check fails closed: it reports not connected unless it can confirm.
  • Update from your agent dashboard.
21 November 2026

Your own voice is now $3 per million characters

  • Plain speech in your own voice dropped from $6 to $3 per million characters, the same price as catalog voices. Every feature is still included.
  • Own voices now run on our Pocket clone engine, on our own servers. Speech in your own voice is returned as a finished mp3 clip.
  • Speech with Polly-compatible calls, including speech marks, is faster and reads phone numbers, order numbers, street numbers, times, dates and titles such as Ms. and Mrs. more accurately.
21 November 2026

100,000 free characters every month, repeat lines free, faster voices, smarter agents

  • Every account gets 100,000 characters of speech free each calendar month, renewing on the 1st. No credit card.
  • The same text in the same voice again returns the stored clip and is not billed or counted (send cache set to off to generate fresh).
  • Speed: Phonon voices 44 ms and Pocket voices 90 ms to the first audible sound (21 November 2026 test, client in AWS us-east-1). Details on the latency page and the streaming guide.
  • Kokoro catalog voices were removed. The catalog is now Phonon and Pocket voices.
  • Agent 0.46.1: speech made by your pay-once agent is not billed per character, it learns repeated answers and keeps the best wording, and it forgets lines and answers nobody used for 90 days. Update from your agent dashboard.
  • MCP server listing updated (0.2.0) with tool titles, annotations and argument descriptions. Python SDK 0.5.0 and JavaScript SDK 0.4.0.
21 November 2026

Speech marks: word times measured from the audio, Amazon-format marks

  • Word times in speech marks ("marks": true) are now measured from the audio instead of spread by word length. On 39 real clips cut into single sentences, the average word-start error against a forced aligner dropped from about 250 ms to about 70 ms, and eight in ten words land within 100 ms.
  • On our own server the word times come from a speech-recognition aligner: two independent aligners agree on word starts to about 15 ms on average. The response field precision.words says which method made the times: model_aligned, aligned or estimated.
  • Amazon Polly drop-in: a plain boto3 client now gets Amazon-format speech marks (OutputFormat json, SpeechMarkTypes word and sentence) with one mark per word of your text and byte offsets into it, text up to 3,000 characters per call, and accented letters and common symbols folded so a café never fails. Polly voice names map to Phonon voices. Not available: viseme and SSML marks. On the same text the character offsets match Polly's on nearly every word.
  • Word times are not engine-exact like Amazon Polly's. Sentence times remain exact.
29 October 2026

Two prices, every feature included; delivery styles discontinued

  • Catalog voices $3 and your own voice $6 per million characters now carry every feature, with nothing to unlock.
  • Audio in the same response: "return":"audio" on POST /v1/speak.
  • Direct streaming for catalog voices from the first sentence (POST /v1/session, paid accounts), billed once per request.
  • Letters such as API, CPU and URL are now read as letters, and long text is split and checked before it is spoken.
  • The $20 styled voice (happy, excited, sorry, calm, serious) is discontinued. A style other than neutral returns style_not_available. Voice fingerprints made for it are no longer kept.
11 October 2026

Every offline voice pack now includes the agent

  • 406 pack listings in 12 languages include the open-source agent. What shipped, what we tested for real and what we did not, and measured speed and price: the full update.
  • Packs page: one row per industry with language and voice selectors.
  • Agent: delivery-status tracking, clearer errors, one Slack channel setting, and 18 languages.
8 October 2026

Speech to text, live streaming and styled voices

  • Speech to text: POST /v1/listen, $0.0015 a minute, 60 free minutes.
  • Styled voices: happy, excited, sorry, calm and serious for your own voice, $20 per million characters.
  • Describe-a-voice chat in the portal.
  • Your own voice can now be used straight from POST /v1/speak.
  • Custom-voice sentences are stored after their second use.
6 October 2026

Free allowance, try-it, MCP, search, agents, status

  • 100,000 free characters per account, shown in the portal.
  • No-signup try-it on the home page and playground.
  • MCP server for Claude, Cursor and other AI tools at /api/mcp.
  • Library search endpoint and offline packs page.
  • Public status page with five-minute checks.
6 October 2026

Live generation in seven languages

  • Any text in English, Spanish, French, Portuguese, Hindi, Japanese and Mandarin Chinese.
  • Friendly errors with a message, a fix and a docs link.
  • Developer portal: keys, test, code, webhooks, usage, AI troubleshooter.
6 October 2026

Site and API launched in development

  • Listening room, voices, pricing calculator, docs and SDKs.
  • Serverless API on CloudFront, Lambda and DynamoDB.