Speech and transcription

Both follow OpenAI's Audio API, so client.audio.speech and client.audio.transcriptions in the OpenAI SDKs work with the base URL changed. The two samples below are files we run against the service before publishing.

Text to speech

# pip install openai
# Text to speech. `voice` is required and its names belong to the model:
# af_heart is one of hexgrad/kokoro-82m's voices.
import os
from openai import OpenAI

client = OpenAI(
    base_url=os.getenv("OCEANTOKEN_BASE_URL", "https://api.oceantoken.ai/v1"),
    api_key=os.environ["OCEANTOKEN_API_KEY"],
)

with client.audio.speech.with_streaming_response.create(
    model="hexgrad/kokoro-82m",
    voice="af_heart",
    input="Welcome aboard. The ferry leaves at nine.",
    response_format="mp3",
) as response:
    response.stream_to_file("speech.mp3")
print("saved speech.mp3")

POST /v1/audio/speech answers with the audio itself.

Field	What it does
model	The ID of a speech model.
input	The text to read.
voice	Required. One of the model's own voice names; see Voices.
response_format	mp3 is the safe choice and works on every model below except where noted. url returns a link instead of the audio (below).
speed, instructions	Pace (0.25 to 4) and tone notes, such as "calm, slow", on models that support them.

Speech is charged per character of input (How you are charged). Send an Idempotency-Key so that a retry after a timeout does not make, and charge, the clip twice (Errors).

Voices

Voice names belong to the model. A request without voice is refused with 400, and a name the model does not know fails. The voices in the second column were tested on the live service (2026-09-26):

Model	A voice that works	Other voices its publisher lists
hexgrad/kokoro-82m	af_heart	af_bella, am_michael, bf_emma, bm_george
deepgram/aura-2	aura-2-thalia-en	aura-2-andromeda-en, aura-2-apollo-en, aura-2-asteria-en, aura-2-helena-en, aura-2-orion-en, aura-2-zeus-en
canopylabs/orpheus-3b-0.1-ft	tara	leah, jess, leo, dan, mia, zac, zoe
sesame/csm-1b	conversational_a	conversational_b
x-ai/grok-voice-tts-1.0	Ara	Rex, Sal, Eve, Leo
microsoft/mai-voice-2	en-US-Jasper:MAI-Voice-1
fish-audio/s1, fish-audio/s2-pro, fish-audio/s2.1-pro, fish-audio/s2.1-pro-free:free	alloy
google/gemini-3.1-flash-tts-preview	Kore (send no response_format; it answers in WAV)	Puck, Charon, Fenrir, Aoede, Leda, Orus, Zephyr

For any other speech model, use the voice names its publisher documents.

Get a link instead of the audio

With "response_format": "url", the answer is JSON with a link to an MP3 instead of the audio bytes, which is easier to pass on to a browser or another request:

{"url": "https://...", "content_type": "audio/mpeg", "bytes": 48213, "usage": {"input_characters": 41}}

Download it within 24 hours. The clip is also in the console under Media.

Transcription

# pip install openai
# Audio file to text. This transcribes speech.mp3, the file the text-to-speech
# sample writes; use any mp3, m4a, wav or webm of your own.
import os
from openai import OpenAI

client = OpenAI(
    base_url=os.getenv("OCEANTOKEN_BASE_URL", "https://api.oceantoken.ai/v1"),
    api_key=os.environ["OCEANTOKEN_API_KEY"],
)

with open("speech.mp3", "rb") as audio:
    transcript = client.audio.transcriptions.create(model="openai/gpt-4o-mini-transcribe", file=audio)
print(transcript.text)

POST /v1/audio/transcriptions takes a multipart upload:

Field	What it does
file	The audio: mp3, m4a, wav, webm and other common formats.
model	The ID of a transcription model.
language	The spoken language as an ISO-639-1 code, such as en or zh. Optional, and improves accuracy when you know it.
prompt	Names and spellings to expect. Optional.
response_format	json (default) returns {"text": "..."}; text the plain text; verbose_json adds segments with timestamps; srt and vtt return subtitles. Models differ in which they offer.

The same with curl:

curl https://api.oceantoken.ai/v1/audio/transcriptions \
  -H "Authorization: Bearer $OCEANTOKEN_API_KEY" \
  -F model="openai/gpt-4o-mini-transcribe" \
  -F file=@meeting.mp3

• The request is at most 10 MB (Limits). Split longer recordings, or compress them: a 10 MB MP3 at 64 kbit/s holds about 20 minutes.
• Transcription is charged per minute of audio (How you are charged).
• The file must be uploaded; a URL in place of the file is not fetched.

Next

• Video: put the voice-over on a clip.
• Prices for every speech and transcription model are on oceantoken.ai.