Skip to content

Python API reference


pip install -U lmnt will install the latest LMNT SDK.


The Speech class is your primary touch-point with the LMNT API. Import it into your module with from lmnt.api import Speech.

When you're done with the Speech instance, make sure to clean up by calling its close() method. Alternatively, you can use this class as an async context manager, which will call close() for you:

async with Speech(API_KEY) as speech:
  voices = await speech.list_voices()

What's new

Date Description
Sep 15, 2023 synthesize now accepts a length parameter to adjust audio output duration to the specified length



Constructing a Speech object requires an API key. Create an API key by visiting your account page in our speech playground and signing up for a (free) plan.


async close()

Releases resources associated with this instance.


async list_voices()

Returns the voices available for use in speech synthesis calls.

Return value

A dictionary of dictionaries (Dict[str, Dict[str, str]]) that describe the available voices. Here's a sample object:

  "shanti": {
    "name": "Shanti",
    "gender": "female",
    "imageUrl": ""




  • The keys of the dictionary (e.g. "shanti") are used to specify the voice in the synthesize speech request below.
  • Some voices may also have an id field; please ignore that field, it will be removed soon.


async synthesize(text, voice, **kwargs)

Synthesizes speech for a supplied text string. Returns binary audio data in one of the supported audio formats.


  • text: the text to synthesize
  • voice: which voice to render; id is found using the list_voices call
  • **kwargs
    • format (optional): aac, mp3, wav; defaults to wav (24kHz 16-bit mono)
    • speed (optional): floating point value between 0.25 (slow) and 2.0 (fast); defaults to 1.0
    • seed (optional): random seed used to specify a different take; defaults to 0
    • durations: (optional) whether to include word durations detail; defaults to false
    • 🆕 length (optional): produce speech of this length in seconds; maximum 300.0 (5 minutes)

Return value

Unless otherwise specified below, the return value is a binary string containing the synthesized audio file.

Duration details

If durations=True is specified, the response will be a dictionary with the following keys:

  • durations: a list of dictionaries with keys as below
  • audio: the binary audio data containing the synthesized audio file

Each durations entry is a dictionary with the following keys:

  • phonemes: a list of the phonemes comprising the word
  • phoneme_durations: a list of the durations of each phoneme
  • start: the starting duration of the word
  • duration: the overall duration of the word

Duration units are given in sample counts at 24kHz.


  • The mp3 bitrate is 96kbps.
  • 🆕 The length parameter specifies how long you want the output speech to be. We will automatically speed up / slow down the speech as needed to fit this length.


async synthesize_streaming(voice)

Creates a new, full-duplex streaming session. You can use the returned session object to concurrently stream text content to the server and receive speech data from the server.


  • voice: which voice to render; id is found using the list_voices call

Return value

A StreamingSynthesisConnection instance, which you can use to stream data.


This class represents a full-duplex streaming connection with the server. The expected use is to call append_text as text is produced and to iterate over the object to read audio. Make sure to call finish() when you're done submitting the entire text snippet.


async append_text(text)

Sends additional text to synthesize to the server. The text can be split at any point. For example, the two snippets below are semantically equivalent:

await conn.append_text('This is a test of ')
await conn.append_text('the emergency broadcast system.')
await speech.append_text('This is a test of the eme')
await speech.append_text('rgency broadcast system.')

__aiter__ / __anext__

async __aiter__()

async __anext__()

Iterates over audio data from the server. See the notes below for details on the audio format. Here's a short snippet that shows how to iterate over the data:

conn = await speech.synthesize_streaming('mara-wilson')
async for msg in conn:
  if msg.type == WSMsgType.CLOSE:
    break  # contains a binary string with the audio data


  • text: some or all of the text to synthesize


  • audio is returned as a 96kbps mono MP3 stream with a sampling rate of 24kHz


async finish()

Call this function when you've written all the text you're expecting to submit. It will flush any remaining data on the server and return the last chunks of audio as described above.

Sample code

Standard synthesis

import asyncio
from lmnt.api import Speech

LMNT_API_KEY = ... # Your API key here
TEXT = 'This is a test of the emergency broadcast system.'

async def main():
  async with Speech(LMNT_API_KEY) as speech:
    voices = await speech.list_voices()
    first_voice = list(voices.keys())[0]
    wav_file = await speech.synthesize(TEXT, first_voice)
    with open('output.wav', 'wb') as f:

if __name__ == '__main__':

Streaming synthesis + ChatGPT

import asyncio
import openai

from argparse import ArgumentParser
from aiohttp import WSMsgType
from lmnt.api import Speech

LMNT_API_KEY = ... # Your LMNT API key here
openai.api_key = ...  # Your OpenAI API key here
MODEL = 'gpt-3.5-turbo'
DEFAULT_PROMPT = "Tell me a story about the people of Foundation."

async def reader_task(conn):
  """Streams audio data from the server and writes it to `output.raw`."""
  with open(f'output.raw', 'wb') as f:
    async for msg in conn:
      if msg.type == WSMsgType.CLOSE:

async def writer_task(conn, prompt):
  """Streams text from ChatGPT to LMNT."""
  response = await openai.ChatCompletion.acreate(
    messages=[{'role': 'user', 'content': prompt}],

  async for chunk in response:
    if 'choices' not in chunk:
    choice = chunk['choices'][0]
    if 'delta' not in choice or 'content' not in choice['delta']:
    await conn.append_text(choice['delta']['content'])
    print(choice['delta']['content'], end='', flush=True)

  await conn.finish()

async def main(args):
  s = Speech(LMNT_API_KEY)
  conn = await s.synthesize_streaming('mara-wilson')

  t1 = asyncio.create_task(reader_task(conn))
  t2 = asyncio.create_task(writer_task(conn, args.prompt))

  await t1
  await t2
  await s.close()

if __name__ == '__main__':
  parser = ArgumentParser()
  parser.add_argument('prompt', default=DEFAULT_PROMPT, nargs='?')