Apppublic

ecohash-ai/tts-studio

sourceHugging Faceapache-2.0updated 10d agoView on Hugging Face

0likes

App README

Text-to-Speech Studio

Synthesize speech with Kokoro-82M or Qwen3-TTS, switch voices, and see the measured latency, real-time factor and exact per-request cost for every generation.

Both models are served by EcoHash on our own NVIDIA RTX PRO 6000 Blackwell (96 GB) fleet — not resold from a third party — through one OpenAI-compatible endpoint at https://api.ecohash.com/v1.

The two models

Kokoro-82M is an 82-million-parameter model designed for latency. It routinely synthesizes faster than real time, which makes it the right default for interactive voice agents and streaming playback. $1.00 per 1M tokens, about $0.006 per minute of audio.

Qwen3-TTS covers ten major languages with richer prosody, at $2.00 per 1M tokens (about $0.012 per audio minute). Use it where naturalness matters more than the last hundred milliseconds.

Reproducing it

python

from openai import OpenAI

client = OpenAI(api_key="eco_YOUR_KEY", base_url="https://api.ecohash.com/v1")

speech = client.audio.speech.create(
    model="kokoro-82m",        # or "qwen3-tts"
    voice="af_heart",
    input="Your text here",
    response_format="wav",
)
speech.write_to_file("out.wav")

Notes

—Use your own API key. The demo runs on EcoHash's key with 5 free runs per visitor per day. Paste your own key into the panel at the top to lift that limit — it is held for your browser session only, never stored and never logged. New accounts get $1 of free credit, which is thousands of requests on these models.
—Text is capped at 800 characters to keep the shared demo responsive; that cap applies whichever key is used.
—Reported latency includes the network round-trip from this Space, so absolute figures are conservative.
—Cost is derived from the published price list and the measured duration of the audio actually returned.

Links

Model catalog · Pricing · Documentation · GPU instances from $1.89/GPU-hour · Get an API key — $1 free credit