Reviewed September 20, 2026

Open-source ElevenLabs alternatives, compared.

Choose by the recording you need: economical narration, expressive speech, multilingual output or local inference. These are task-based candidates, with samples where available and explicit licensing differences.

Why people look for an alternative

Kokoro costs $0.02 per 1,000 characters in OpenSpeech, with a $0.01 minimum per recording and whole-cent rounding. At exactly 1,000 characters that is 60% below ElevenLabs Flash/Turbo’s $0.05 direct API rate. Paragraph projects sum individual paragraph prices. This compares usage charges, not feature parity or voice quality. Official ElevenLabs API prices, checked September 20, 2026. Create an affordable voiceover →

Hear the same-script benchmark and inspect the measurements →

Six candidates worth evaluating

01

Kokoro-82M

Economical narration

Eight English voices in our paid studio. The paragraph editor combines recordings into an MP3 or WAV. No voice cloning. Compare the existing samples with your own script.

Apache-2.0English, Japanese, Chinese, Spanish, French, Hindi, Italian, PortuguesePreset voicesSources, samples & setup
02

Qwen3-TTS

Multilingual and controlled speech

The upstream family offers presets, design and cloning modes. Our current hosted integration exposes an English preset; not all upstream capabilities are available in our studio.

Apache-2.0Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, ItalianVoice cloningSources, samples & setup
03

Chatterbox Turbo

Expressive short speech

An MIT-licensed speech model with expressive controls. Our hosted preset accepts up to 500 characters. Its generation price is higher than Kokoro.

MITEnglishVoice cloningSources, samples & setup
04

VoxCPM2

Multilingual voice design

A 2B model with Apache-licensed code and weights. Evaluate the official checkpoint on your languages and scripts. Not currently hosted by OpenSpeech.

Apache-2.0Arabic, MY, Chinese, DA, Dutch, English, FI, French, German, EL, HE, Hindi, Indonesian, Italian, Japanese, KM, Korean, LO, MS, NO, Polish, Portuguese, Russian, Spanish, SW, SV, TL, Thai, Turkish, VietnameseVoice cloningSources, samples & setup
05

Soprano 1.1

Compact English inference

An 80M English model supporting CPU and GPU inference. No cloning. We measured local generation; cloud economics and blind quality comparisons remain unverified.

Apache-2.0EnglishPreset voicesSources, samples & setup
06

Pocket TTS

CPU speech generation

The original English checkpoint has CC-BY-4.0 weights. Reference-voice licenses need separate review. Newer language checkpoints require their own checks.

MIT (code) · CC-BY-4.0 (weights)English, French, German, Portuguese, Italian, SpanishVoice cloningSources, samples & setup

ElevenLabs alternatives FAQ

Which open model should I try first?

Kokoro is our starting point for economical English narration. Evaluate Qwen3-TTS and Chatterbox for different speech styles. We have not established an overall quality winner.

Are these models free for commercial use?

Review code, weights, tokenizers and reference voices separately. Permissive model terms still have conditions. Fish S2 Pro requires a separate commercial license; XTTS and F5 weights also have restrictions. Higgs weights use a custom community agreement.

Is OpenSpeech cheaper than ElevenLabs?

At exactly 1,000 characters, OpenSpeech Kokoro costs $0.02 versus the current $0.05 ElevenLabs Flash/Turbo direct API rate. This compares generation charges, not quality or feature parity. A $0.01 recording minimum and whole-cent rounding affect shorter scripts; credit packs start at $5.

What does self-hosting cost?

Model weights can be free to use while compute, storage, idle capacity and maintenance still cost money. Benchmark your actual deployment and include failed attempts. Local CPU timing is not a measured cloud bill.

How should I compare quality?

Use the same original scripts and record the checkpoint, voice and settings. Listen for missing words, repetition, pronunciation and clean endings. Existing samples are auditions; missing samples and incomplete reviews are labeled.