Voice Cloning Streaming API

The Voice Cloning Streaming API accepts reference audio and text to synthesize in one multipart request, then returns an audio binary stream. The reference transcript is optional.

API overview

Item	Value
Base URL	https://api.kitschlabs.com
Authentication	xi-api-key: <API_KEY>
Method	POST
Endpoint	/v1/voice-cloning/text-to-speech/stream
Content-Type	multipart/form-data
Default output	mp3_44100_128 (audio/mpeg)
Additional output	wav_24000 (audio/wav)

How do I generate speech?

When provided, reference_text must match the reference audio exactly. If it is omitted or blank, the service automatically transcribes the reference audio.

curl --request POST \
  "https://api.kitschlabs.com/v1/voice-cloning/text-to-speech/stream?output_format=wav_24000" \
  --header "xi-api-key: <API_KEY>" \
  --form "text=This is the text to synthesize." \
  --form "reference_text=This is the exact reference transcript." \
  --form "reference_audio=@reference.wav" \
  --form "language_code=en" \
  --output speech.wav

Which fields do I send?

Field	Location	Type	Required	Description
text	form	string	Yes	Text to synthesize
reference_text	form	string	No	Exact transcript of the reference audio. Automatically transcribed when omitted
reference_audio	form	file	Yes	WAV, MP3, FLAC, or OGG file
language_code	form	string	No	One of en, ko, ja, or zh
instruct	form	string	No	Natural-language instructions for the speaking style
output_format	query	string	No	mp3_44100_128 or wav_24000

Reference audio can be up to 25 MiB and 20 seconds. When providing reference_text, include only the words that can be heard. Do not include timestamps, speaker names, subtitle markers, or translations.

A successful request returns audio binary in the selected output_format. See the Voice Cloning API for required values, reference audio guidance, and error handling.