Paxa Labs for Dify

Speech models from Paxa Labs for Dify apps: text to
speech and speech to text for Thai, English, and Mandarin Chinese.

• Source: https://github.com/paxalabs/dify-plugin-paxa
• Author: Paxa Labs, [email protected]
• Plugin type: model provider (text to speech, speech to text)

Models

Setup

1. Create an API key at https://paxalabs.com/app/keys. New accounts include
   free credits.
2. In Dify, install the Paxa Labs plugin from the Marketplace.
3. Go to Settings > Model Provider, open Paxa Labs, and enter your API
   key. Dify checks the key with a free call before it saves it.
4. In the system model settings on the same page, choose
   as the text-to-speech model and
   as the speech-to-text model.

Usage

• Read answers aloud. In an app's Features, turn on Text to
  Speech, then pick a language and a voice. Dify speaks answers with the
  voice you picked, as each answer streams in when auto-play is on, or when a
  user presses the speaker button.
• Talk to an app. In Features, turn on Speech to Text. Users can
  then speak through the microphone button, and Dify transcribes the
  recording with Paxa STT Lite.
• Voices. The voice list comes from the live Paxa voice catalog, so new
  voices appear without a plugin update. Foi Thong (female) and Massaman
  (male) are listed first. On Paxa TTS Flash v2 every voice suits all three
  languages. On Paxa TTS Flash, the language setting narrows the list to the
  voices designed for that language.
• Direction tags. On Paxa TTS Flash v2, a tag in the text changes the
  delivery from where it stands, for example
  The tag list is at
  https://paxalabs.com/docs/directions.

Credentials

The plugin needs one credential: a Paxa API key () from
https://paxalabs.com/app/keys. The key is sent only to the Paxa API, in the
header. A spending cap set on the key in the Paxa dashboard
applies to Dify's calls as well.

Connection requirements

The plugin connects to one host, , over HTTPS on port 443:

Every request carries an header.
A self-hosted Dify needs outbound HTTPS access to from its
plugin daemon. The plugin uses no other host, webhook, or callback.

Pricing and limits

Usage is billed to your Paxa account in credits (1,000 credits = US$1). At the
time of this release:

• Text to speech costs 10 credits per 1,000 characters on both models.
• Speech to text costs 500 credits per hour of audio, billed by the
  recording's exact length.

Current rates are at https://paxalabs.com/pricing.

• A speech request takes up to 5,000 characters. The plugin sends longer
  text as consecutive requests, split at sentence ends or spaces. A direction
  tag applies within its own part.
• A recording can be up to 60 minutes long and 25 MB in size, as MP3, WAV,
  FLAC, Ogg, M4A, AAC, or WebM.
• Request rate and concurrency follow your Paxa plan.

Privacy

See PRIVACY.md for what the plugin sends and how Paxa Labs
handles it.

Support

• API documentation: https://paxalabs.com/docs
• Issues: https://github.com/paxalabs/dify-plugin-paxa/issues
• Email: [email protected]

Development

The last command uses the Dify plugin CLI.
To debug against a Dify instance, copy to , set the
debug key from the instance's plugin page, and run .

refreshes the voice lists bundled in
from the public catalog. The plugin reads the live
catalog at runtime; the bundled lists serve Dify screens that read voices
from the model schema.

License

MIT