Speech API · Whisper large-v3-turbo + Kokoro + Chatterbox · EU-processed
Speech-to-text and text-to-speech API, processed in the EU
The FastInfra speech API transcribes and translates audio with Whisper large-v3-turbo at $0.003 per audio minute, half the price of OpenAI whisper-1, and speaks text with Kokoro-82M and Chatterbox Multilingual from $0.30 per million characters. All three models run on a FastInfra GPU in Romania, so audio stays in the EU and is not stored after the response or used for training. The endpoints match the OpenAI audio API, so the official SDKs work after changing the base URL, and the page shows live request counts and the pod's status.
Transcribe, translate and speak through OpenAI-compatible endpoints on a FastInfra GPU in the EU. $0.003 per audio minute, $0.30–$0.50 per million characters of speech. Drop-in for the OpenAI SDK: change the base URL and keep your code.
- EU-processed (Romania)
- No training on your audio
- Audio not stored after the response
- OpenAI SDK drop-in
100.0% uptime over 299 probes in 24 h · median probe 59 ms
Measured, not estimated
Counted from the billing roll-up, not from the pod's health counter (which resets on restart). Refreshed every ten minutes.
Reliability: All three models share one RTX 4090 pod in RunPod EU-RO-1 today; the status page shows its 24-hour uptime, and it did restart once during our last audit. A second replica is the prerequisite for printing an uptime figure on this page.
Transcribe and translate
Whisper large-v3-turbo (INT8) with word-level verbose_json, language detection, 2-hour / 25 MB files and a realtime WebSocket. $0.003 per audio minute, half of OpenAI whisper-1.
Speak in 23 languages
Kokoro-82M for fast preset voices in 8 languages at $0.30 per million characters; Chatterbox Multilingual for expressive speech and voice cloning in 23 languages at $0.50. Streaming starts after the first sentence.
Processed in the EU
Audio is sent to our GPU in Romania and the text returned; nothing is written to disk after the response, nothing is used for training. See the data-handling policy linked below.
How speech is priced
| What | Price | Token equivalent | Notes |
|---|---|---|---|
| Transcription / translation | $0.003 per audio minute | 6,000 output tokens per minute at $0.50 / 1M | Billed on audio duration, rounded up to the second; realtime sessions bill wall-clock |
| Kokoro text-to-speech | $0.30 per 1M characters | 1 character = 1 output token; $0.0003 per 1,000 characters | Preset voices, 8 languages, fastest |
| Chatterbox text-to-speech | $0.50 per 1M characters | 1 character = 1 output token; $0.0005 per 1,000 characters | 23 languages, zero-shot voice cloning |
Worked example
- 1-hour meeting recording$0.18
- 1,000 hours of call audio$180.00
- A 2,000-character article read by Kokoro$0.0006
- A 20,000-word audiobook chapter (~120k characters) by Chatterbox$0.06
Prices are USD, prepaid credits, no platform fee. Transcription was $0.006 per minute (OpenAI whisper-1 parity) until November 2026; the cut to $0.003 is permanent, not a promotion.
Try it in the browser
Record up to 30 seconds or type a sentence. Demo calls use a short-lived page token and a small per-visitor budget; they never touch a real API key.
Record up to 30 seconds in any language, or pick a file. The transcript comes back with the detected language and what the same call would have cost through the API.
Same as transcribe, but the output is always English. Speak in Hindi, German, Spanish or anything else Whisper knows.
Type up to 200 characters, pick a voice, listen. Kokoro answers in well under a second; Chatterbox takes a moment longer and speaks 23 languages.
Code
# transcribe
curl https://api.fastinfra.ai/v1/audio/transcriptions \
-H "Authorization: Bearer $FASTINFRA_API_KEY" \
-F file=@meeting.mp3 \
-F model=whisper-large-v3-turbo \
-F response_format=verbose_json
# speak
curl https://api.fastinfra.ai/v1/audio/speech \
-H "Authorization: Bearer $FASTINFRA_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model": "kokoro", "voice": "af_heart", "input": "Hello from FastInfra."}' \
-o hello.wav
from openai import OpenAI
client = OpenAI(base_url="https://api.fastinfra.ai/v1", api_key=API_KEY)
text = client.audio.transcriptions.create(
model="whisper-large-v3-turbo",
file=open("meeting.mp3", "rb"),
response_format="verbose_json",
)
speech = client.audio.speech.create(
model="chatterbox", voice="default", input="Bonjour et bienvenue.",
extra_body={"language": "fr"},
)
speech.write_to_file("bonjour.wav")
import OpenAI from "openai";
import fs from "node:fs";
const client = new OpenAI({ baseURL: "https://api.fastinfra.ai/v1", apiKey: process.env.FASTINFRA_API_KEY });
const t = await client.audio.transcriptions.create({
model: "whisper-large-v3-turbo",
file: fs.createReadStream("meeting.mp3"),
});
console.log(t.text);
const wav = await client.audio.speech.create({ model: "kokoro", voice: "af_heart", input: "Hello from FastInfra." });
fs.writeFileSync("hello.wav", Buffer.from(await wav.arrayBuffer()));
// Stream microphone PCM16 frames; receive partial and final transcripts.
const ws = new WebSocket("wss://api.fastinfra.ai/v1/realtime/transcription?api_key=" + API_KEY);
ws.onmessage = (e) => console.log(JSON.parse(e.data));
ws.onopen = () => ws.send(JSON.stringify({ type: "session.update", language: "en" }));
// ws.send(pcm16Frame) for each 100 ms of 16 kHz mono audio
Volume
Pay as you go
Any volume
$0.003 / min
Prepaid credits, no commitment, no minimum.
Committed volume
From 10,000 audio hours / month
Quoted
Send your monthly hours through the contact form; we quote a committed rate from the support queue.
Dedicated pod
Your own RTX 4090
$1.20 / hr
Rent an RTX 4090 on GPU Cloud and run the same Whisper + Kokoro image: unlimited minutes for the hourly rate, under $0.003 per audio minute at 8 concurrent streams.
Committed-volume requests go through the contact form and land in the same support queue as everything else. Contact form.
Compared with
| Provider | Published price | Notes |
|---|---|---|
| OpenAI whisper-1 | $0.006 / min | Same model family; twice the FastInfra price, US-processed |
| OpenAI gpt-4o-mini-transcribe | $0.003 / min | Same price as FastInfra; no TTS bundle, US-processed |
| Deepgram Nova-3 | $0.0043 / min | Batch, monolingual |
| AssemblyAI Universal-2 | $0.0025 / min | Cheaper per minute; no EU-hosted option on pay-as-you-go |
| Hosted Whisper discounters | $0.0007–0.0018 / min | Cheapest raw Whisper; US-processed, no TTS, rate-limited free tiers |
Competitor prices are their published list prices in November 2026 and may change. We are not the cheapest Whisper on the internet; we are the cheapest one that is processed in the EU and ships with TTS on the same key.
Voices
-
af_heartHeart · American English · female kokoro -
am_michaelMichael · American English · male kokoro -
bf_emmaEmma · British English · female kokoro -
ef_doraDora · Spanish · female kokoro -
ff_siwisSiwis · French · female kokoro -
hf_alphaAlpha · Hindi · female kokoro -
if_saraSara · Italian · female kokoro -
pf_doraDora · Brazilian Portuguese · female kokoro -
zf_xiaobeiXiaobei · Mandarin · female kokoro -
defaultChatterbox built-in voice · 23 languages chatterbox
Listen renders a short sample through the demo above and counts toward its budget.
Limits and formats
- Files up to 25 MB and 2 hours per request; mp3, wav, m4a, webm, ogg, flac.
- Realtime transcription over WebSocket bills wall-clock session time at the per-minute rate.
- Text-to-speech input up to 4,096 characters per request; use the async jobs endpoint for long passages.
- Output is 24 kHz mono WAV (or raw PCM16); set stream=true to start playback after the first sentence.
- 8 concurrent transcription streams per pod today; 429 with Retry-After when the queue is full.
Where it runs and what we keep: see regions and data handling.
Frequently asked questions
How much does transcription cost?
$0.003 per audio minute, billed on the duration of the file rounded up to the second, which is $0.18 per hour of audio. That is half of OpenAI whisper-1 and below Deepgram and Google batch rates; realtime WebSocket sessions bill wall-clock time at the same rate.
Which speech models are available?
Whisper large-v3-turbo for transcription and translation to English, Kokoro-82M for fast preset voices in eight languages at $0.30 per million characters, and Chatterbox Multilingual for expressive speech and voice cloning in 23 languages at $0.50 per million characters.
Is the speech API compatible with the OpenAI SDK?
Yes. Point the OpenAI client at the FastInfra base URL and call audio.transcriptions.create, audio.translations.create or audio.speech.create as usual. Model aliases whisper-1, tts-1 and tts-1-hd map to the self-hosted models so existing code runs unchanged.
Where is my audio processed and is it stored?
On a FastInfra GPU in RunPod EU-RO-1, Romania. The audio is held in memory for the request, the text is returned, and nothing is written to disk afterwards unless you opt a key into body capture. Nothing is used for training, and the policy is written out on the security page.
What are the speech API limits?
Uploads up to 25 MB and 2 hours per request in mp3, wav, m4a, webm, ogg or flac; text-to-speech input up to 4,096 characters per request with an async jobs endpoint for longer passages; eight concurrent transcription streams per pod with a 429 and Retry-After when the queue is full.
Reviewed by Ashish Kumar, Head of Product · Updated November 17, 2026. Prices and counts on this page come from the live catalog.
Half the price of whisper-1, processed in the EU
Create a free account, point the OpenAI SDK at https://api.fastinfra.ai/v1 and transcribe your first hour for $0.18.