Which open-source TTS should I try?

Start with the three picks below. Hear them. Pick one. We’ll get into the rest later.

Use your text in the paid studio →Explore newly added models

Start here

Editorial picks · try the community arena

Everyday narration

Kokoro-82M

Compact English speech with eight hosted voices and paragraph editing from $0.02 per 1,000 characters.

Voice

Multilingual research

Qwen3-TTS

An open speech family with multilingual capabilities. Check each checkpoint and the narrower hosted preset before choosing.

VoiceDefault

Expressive short speech

Chatterbox Turbo

A compact expressive model to evaluate for short clips. The hosted preset accepts up to 500 characters per recording.

VoiceAndy

Side-by-side

Hear the leaders back-to-back

Same three scripts, three top models, one screen. The fastest way to decide what you actually want.

Kokoro-82MQwen3-TTSChatterbox Turbo

Compare these three

New to the directory

See all recent additions →

New releases get their own pages and recordings. Start with these additions, then compare them with the established models.

Added 2026-09-14 · Listen below

Chatterbox Turbo

Andy

English voice cloning and expressive speech with a smaller model.

Permissive model licenseCloning350M

Added 2026-09-14 · Listen below

Pocket TTS

Alba

CPU speech generation, streaming, and voice cloning in a small model.

Permissive model licenseCloningStreamingCPU100M

Added 2026-09-14 · Samples pending

MOSS-TTS v1.5

Default

Long-form multilingual narration with cloning and explicit pause controls.

No samples yet

Permissive model licenseCloning8B

The three scripts (every voice reads these)

Neutral

“The quick brown fox jumps over the lazy dog near the riverbank.”

Baseline naturalness — no emotion, no tricky words

Emotional

“I can't believe you actually did it. This is incredible!”

Prosody and excitement — tests expressiveness

Numbers & Dates

“On March 14th, 2025, the team raised $4.2 million at a 38% margin.”

Number, date, and symbol pronunciation — the common failure mode

Browse by need

Not sure what you need yet?

Each shelf below is hand-curated. Scroll horizontally to see all picks. Add anything to the comparison bar to hear them side-by-side.

Best for voice cloning

Clone a voice from a short reference clip.

Chatterbox

Default

English voice cloning with emotion-exaggeration controls

Permissive model licenseCloning500M· 8GB

XTTS v2

Reference clone

Voice cloning from a short reference clip across 17 languages

Noncommercial / separate licenseCloningStreaming750M· 6GB

Fish Speech 1.5

Reference clone

Earlier Fish Speech voice-cloning model; compare with S2 Pro

Noncommercial / separate licenseCloningStreaming0.5B· 12GB

Bark

Generates laughter, sighs, and music — the chaotic one

Permissive model licenseCloning1.5B· 12GB

F5-TTS

Reference clone

Flow-matching TTS, fast zero-shot voice cloning

Noncommercial / separate licenseCloning335M· 8GB

Dia

Ultra-realistic two-speaker dialogue, sounds like a podcast

Permissive model licenseCloningStreaming1.6B· 10GB

CosyVoice 2

Default

Alibaba's real-time TTS, 150ms streaming latency

Permissive model licenseCloningStreaming500M· 6GB

IndexTTS 2

Default

Bilibili's TTS with multi-speaker and precise emotion control

Commercial conditionsCloningStreaming1B· 8GB

Best for real-time

Stream tokens fast enough for voice agents.

KittenTTS

Default

15M params, under 25MB, runs anywhere — even your phone

Permissive model licenseStreamingCPU15M

MeloTTS

Default

Multilingual, real-time on CPU

Permissive model licenseStreamingCPU52M

Parler-TTS

Describe the voice in natural language ('soft female, fast, clear')

Permissive model licenseStreaming880M· 6GB

Qwen3-TTS

Default

Alibaba's flagship, 97ms latency, 10 languages

Permissive model licenseCloningStreaming1.7B· 8GB

Fish Speech 1.5

Reference clone

Earlier Fish Speech voice-cloning model; compare with S2 Pro

Noncommercial / separate licenseCloningStreaming0.5B· 12GB

CosyVoice 2

Default

Alibaba's real-time TTS, 150ms streaming latency

Permissive model licenseCloningStreaming500M· 6GB

Spark-TTS

Default

Built on Qwen2.5, zero-shot voice cloning + style control

Noncommercial / separate licenseCloningStreaming500M· 8GB

IndexTTS 2

Default

Bilibili's TTS with multi-speaker and precise emotion control

Commercial conditionsCloningStreaming1B· 8GB

Best for CPU

No GPU? No problem. These run on a laptop.

Kokoro-82M

82M parameters, preset voices and CPU inference

Permissive model licenseStreamingCPU82M

Piper

LJ Speech

Tiny, fast, runs offline on a Raspberry Pi

Commercial conditionsStreamingCPUVoice-dependent

KittenTTS

Default

15M params, under 25MB, runs anywhere — even your phone

Permissive model licenseStreamingCPU15M

MeloTTS

Default

Multilingual, real-time on CPU

Permissive model licenseStreamingCPU52M

NeuTTS Air

Reference-based

On-device-class TTS with instant cloning from ~3 seconds

Commercial conditionsCloningStreamingCPU748M

Added 2026-09-14 · Listen below