Now live New models for transcription and text-to-speech. AssemblyAI-STT-3.7 AssemblyAI-TTS-1
Insights & Use Cases

Best free TTS APIs and open-source text-to-speech models (2026)

Kokoro, Piper, Qwen3-TTS, Chatterbox and more: open-source TTS models compared by license, maintenance and real-time fit, plus free TTS API tiers.

Abstract green cone illustration

Written by

Kelsey Foster

Published on

18 November 2026

Open-source text-to-speech has had a busy two years. Hugging Face counted more than 8,000 TTS models on its Hub when it launched its Open TTS Leaderboard on Sep 30, 2026. Most "best open source TTS" lists still rank those models on how they sound, then skip the two questions that decide whether you can ship one: what the license lets you do, and whether anyone still maintains it.

Those questions aren't academic. Several popular "open-source" models have open code but non-commercial weights. Piper moved to a GPL-3.0 codebase. Chatterbox, one of the most-recommended cloning models of 2025, was retired in August 2026. The original Coqui TTS repo hasn't had a commit since February 2024, and coqui.ai no longer loads.

This guide compares open-source TTS models by code license, weights license, commercial use, maintenance and real-time fit. Every license and commit date was checked on Nov 12, 2026. It also covers free TTS APIs and free tiers, what self-hosting costs once you're in production, and where a managed API fits. If you're looking for the speech-to-text side, see our companion guide to free speech-to-text APIs and open-source engines.

Quick answer

  • Small, permissive and fast: Kokoro-82M. It's 82 million parameters, with Apache-2.0 code and weights, and it runs comfortably on modest hardware.
  • Voice cloning under a permissive license, still maintained: Qwen3-TTS. Apache-2.0 code and weights, 10 languages, 3-second cloning and streaming output.
  • Offline and edge devices: Piper, with one caveat. Development moved to a GPL-3.0-or-later codebase, and each voice carries its own dataset license.
  • A voice agent in production: a managed streaming TTS API, unless you're ready to run GPU serving, streaming input, interruption handling and telephony audio yourself.

What "free" means in text-to-speech

"Free text to speech" covers three different things. They come with different costs, limits and legal terms.

Question Free tier of a TTS API Open-source / open-weight model Browser and OS voices
What's free? A monthly character or credit allowance, then pay as you go The code and weights. You pay for the hardware. The voices already installed on the user's device
Where does it run? The vendor's cloud Your GPU, CPU or device The user's browser or operating system
What should you read? The pricing page and the terms of the free plan The code license and the weights license, separately Browser support and which voices each platform ships
Main limit Allowance size, rate limits, sometimes commercial-use rights Serving, scaling and maintenance are your job You don't control the voice, and it varies by device
Good for Prototypes and evaluations before you commit Offline apps, privacy-bound workloads, research, fine-tuning Accessibility features and simple read-aloud in web apps

The rest of this post takes them in order: open-source models first, then free APIs and tiers.

Open-source TTS models compared

The table below separates the code license (the inference and training code on GitHub) from the weights license (the trained model you download, usually from Hugging Face). They often differ, and the weights license is the one that decides commercial use. "Last commit" is the newest commit on the default branch. A date more than a year old usually means nobody is fixing bugs.

Model Code license Weights license Commercial use? Last commit (checked Nov 12, 2026) Languages Voice cloning Streaming
Kokoro-82M Apache-2.0 Apache-2.0 Yes Aug 6, 2025 8 (54 voices) No Segment by segment
Piper (OHF-Voice/piper1-gpl) GPL-3.0-or-later Per voice. The project grants none. Engine under GPL terms; check each voice's datasets Nov 11, 2026 Dozens of languages and locales No (train or fine-tune a voice) Not stated
Qwen3-TTS (0.6B, 1.7B) Apache-2.0 Apache-2.0 Yes Mar 17, 2026 10 Yes (3-second clone) Yes
MOSS-TTS family Apache-2.0 Apache-2.0 Yes Sep 6, 2026 31 (v1.5) Yes Yes (MOSS-TTS-Realtime)
Orpheus (Canopy Labs)* Apache-2.0 Apache-2.0 tag, gated* Yes, see footnote* Dec 5, 2025 English (multilingual research preview) Yes (zero-shot) Yes
Dia 1.6B (Nari Labs) Apache-2.0 Apache-2.0 Yes, with forbidden-use terms Nov 19, 2025 English Yes (audio prompt) Not stated
NeuTTS Air (Neuphonic) NeuTTS Open License 1.0 Apache-2.0, gated Weights yes; repo code only below $5M annual revenue Jul 22, 2026 English Yes (3 seconds of audio) GGUF builds only
VibeVoice-Realtime-0.5B (Microsoft) MIT MIT Licensed, but the README says research use only Oct 8, 2026 English, plus experimental voices in 9 languages Not stated Yes, including streaming text input
Chatterbox (Resemble AI) MIT MIT Yes, but retired: no updates or support Nov 4, 2026 (retirement notice) 23 (Multilingual V3) Yes Not stated
Sesame CSM-1B* Apache-2.0 Apache-2.0 tag, gated* Yes, see footnote* May 27, 2025 English Yes (audio context) Not stated
F5-TTS MIT CC-BY-NC-4.0 No (non-commercial weights) Nov 3, 2026 Varies by checkpoint Yes Not stated
XTTS-v2 (idiap/coqui-ai-TTS fork) MPL-2.0 Coqui Public Model License No (non-commercial weights) Oct 24, 2026 (fork) 17 Yes (6-second clip) Yes

*Orpheus is built on Llama 3.2 3B, and its model card lists meta-llama/Llama-3.2-3B-Instruct as a base model. CSM-1B uses a Llama backbone and needs access to Llama-3.2-1B to run. Both weights carry an Apache-2.0 tag on Hugging Face, but the Llama 3.2 Community License may also apply to models derived from Llama. Read both licenses before you ship either one commercially.

Sources: each project's GitHub repo (license, README and latest commit on the default branch) and Hugging Face model card, checked on Nov 12, 2026. "Not stated" means the project's README doesn't document the capability. This is a summary, not legal advice. Read the license files themselves before you ship.

Hear AssemblyAI-TTS-1 on your own text

Comparing open-source voices against a managed API? Paste your own agent script, with order numbers, dates and names, into the playground and listen before you decide.

Try playground

The open-source TTS models, one by one

Kokoro TTS (Kokoro-82M)

Kokoro is the default answer to "what's a good small open-source TTS model?", and for good reason. It has 82 million parameters and Apache-2.0 code and weights. The model card says it "has been deployed in numerous projects and commercial APIs". Version 1.0 ships 54 voices across 8 languages: American and British English, Japanese, Mandarin, Spanish, French, Hindi, Italian and Brazilian Portuguese.

The trade-offs: there's no voice cloning, and the Python pipeline returns audio one text segment at a time rather than streaming token by token. The GitHub repo is quiet. The last commit was Aug 6, 2025, and the newest PyPI release (kokoro 0.9.4) is from April 2025. It's stable, but don't expect fixes.

Piper TTS (now GPL-3.0)

Piper is the go-to engine for offline and edge text-to-speech: home assistants, Raspberry Pi projects and screen readers. Two things changed that most lists haven't caught.

First, the original rhasspy/piper repository is archived. Development moved to OHF-Voice/piper1-gpl, which is active (last commit Nov 11, 2026) and licensed GPL-3.0-or-later. The PyPI package piper-tts 1.8.0 carries the same license. That matters if you embed Piper in a closed-source product, because the old repo was MIT.

Second, the project's voice licensing page says Piper "makes no license grant of its own over voice models". Each voice's rights come from its training data. The project keeps a separate list of voices whose entire training chain uses commercially usable recordings. Start there if you're shipping a product.

Qwen3-TTS

Qwen3-TTS is the strongest permissive pick for cloning today. The repo and the 0.6B and 1.7B weights are all Apache-2.0. It covers 10 languages: Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish and Italian. The Base models clone a voice from 3 seconds of audio, and the README describes streaming output with end-to-end synthesis latency "as low as 97ms". The last commit was Mar 17, 2026, so it's maintained, but not daily.

Chatterbox TTS (retired)

Chatterbox was one of the most-recommended open-source cloning models of 2025: MIT code and weights, a 350M-parameter Turbo model for English voice agents, and a 0.5B Multilingual V3 model covering 23 languages. Every output carries Resemble AI's PerTh watermark.

It's also finished. The README now opens with a notice that "as of August 2026, Chatterbox is retired research and is no longer maintained". The code stays public under MIT with no support or updates. You can still use it commercially. You just own every bug.

Orpheus

Orpheus is an LLM-style TTS model built on a Llama 3.2 3B backbone, with zero-shot cloning and streaming. Its README quotes about 200 ms streaming latency, "reducible to ~100ms with input streaming". The production models are English, and a multilingual family is out as a research preview. Code and weights are Apache-2.0, the weights are gated on Hugging Face, and the Llama footnote above applies. The last commit was Dec 5, 2025.

Dia

Dia is a 1.6B-parameter model built for dialogue. It generates conversation directly from a transcript, including nonverbal sounds like laughter and coughing, and clones a voice from a 5- to 10-second audio prompt. Code and weights are Apache-2.0. The README says the model "only supports English generation at the moment" and lists forbidden uses, including producing audio that resembles real people without permission. The last commit was Nov 19, 2025, when Nari Labs pointed users to its successor, Dia2: a streaming dialogue model with 1B and 2B checkpoints, also Apache-2.0 and English only, with no commits since Nov 29, 2025.

F5-TTS (non-commercial weights)

F5-TTS is active (last commit Nov 3, 2026) and popular for cloning, which is why its license trips people up. The code is MIT, but the README states that "the pre-trained models are licensed under the CC-BY-NC license" because of the Emilia training data. You can study it and build on the code. You can't ship the released weights in a commercial product.

Coqui TTS and XTTS-v2: what replaced Coqui

Coqui TTS was the open-source TTS toolkit for years. Today the original coqui-ai/TTS repo's last commit is from Feb 10, 2024, and coqui.ai, the site its model license points to, no longer loads.

The maintained successor is the Idiap Research Institute's fork, idiap/coqui-ai-TTS. It's MPL-2.0 and active (last commit Oct 24, 2026), and it publishes to PyPI as coqui-tts (0.28.0). If you're searching for "coqui tts" today, that's the package to install.

The fork fixes the code, not the license on the weights. XTTS-v2, the toolkit's best-known model, clones a voice from a 6-second clip in 17 languages and streams, but its weights remain under the Coqui Public Model License, which "allows only non-commercial use of a machine learning model and its outputs". With the licensor's site offline, there's no obvious place to buy a commercial license.

MOSS-TTS

The MOSS-TTS family from the OpenMOSS team is Apache-2.0 throughout and actively developed (last commit Sep 6, 2026). It's a family rather than one model. The 8B MOSS-TTS v1.5 covers 31 languages with voice cloning. MOSS-TTS-Realtime is built for low-latency voice agents. MOSS-TTS-Nano (about 100M parameters) streams on a CPU. The 8B model is a real GPU commitment, though the project ships llama.cpp and GGUF builds to shrink it.

NeuTTS Air

NeuTTS targets on-device speech with instant cloning from about 3 seconds of audio and watermarked outputs. Read the license carefully. The NeuTTS Air weights are Apache-2.0 (English only, gated on Hugging Face). The repository's LICENSE file, though, is the NeuTTS Open License 1.0, which covers the newer Nano models and limits commercial use to organizations with under $5 million in annual revenue. Streaming works with the GGUF builds only.

VibeVoice

Microsoft's VibeVoice is MIT-licensed, and its 0.5B Realtime model streams text input with about 300 ms to first audible speech, according to the README. Two caveats. Microsoft removed the original long-form VibeVoice-TTS code in September 2025 after finding it "used in ways inconsistent with the stated intent". And the README says it doesn't recommend VibeVoice "in commercial or real-world applications without further testing and development".

Sesame CSM-1B (notable but quiet)

CSM is a conversational speech model that generates speech from text plus audio context. It's Apache-2.0, the weights are gated, and it needs Llama-3.2-1B to run (see the footnote). The README frames it for research and educational use and says non-English performance "likely won't" be good. There have been no commits since May 27, 2025.

Non-commercial or restrictive licenses

These projects publish weights you can download, but the weights license limits commercial use. Some lists call them "open source". Treat them as weights-available.

  • F5-TTS: CC-BY-NC-4.0 weights. Non-commercial.
  • XTTS-v2: Coqui Public Model License. Non-commercial, and the licensor's site is offline.
  • Fish Audio (Fish Speech, S2-Pro, Drama 3): the fish-speech code is under the Fish Audio Community License (updated Oct 13, 2026). It allows commercial use only while your annual revenue is under $500,000, and never as a hosted TTS service. The S2-Pro and Drama 3 weights stay under the Fish Audio Research License. Our Fish Audio speech generation guide walks through what each license allows, and our Fish Audio alternatives roundup covers the options.
  • IndexTTS-2 (bilibili): a custom model license. Services with more than 100 million monthly active users or more than RMB 1 billion in annual revenue need a separate license, and you can't use it to improve other commercial AI models.
  • Higgs Audio (Boson AI): the code is Apache-2.0, but the v2 weights use a community license based on the Meta Llama 3 Community License, and the README places v3 under a "Research and Non-Commercial License".
  • NeuTTS Nano models: NeuTTS Open License 1.0. Commercial use only below $5M in annual revenue.

Legacy engines you'll still see recommended

Older lists still lead with these. Here's where they stand:

  • eSpeak NG: GPL-3.0 and still active (last commit Nov 8, 2026). Robotic, but tiny and fast, and Kokoro uses it as a fallback for some languages.
  • Bark (MIT), StyleTTS2 (MIT), Parler-TTS (Apache-2.0) and MeloTTS (MIT): permissive, but no commits since 2024. Zonos (Apache-2.0) has been quiet since March 2025.
  • MaryTTS and Festival: they still turn up in older "open source TTS" lists. Check each repo's recent activity before you build on them.

For a neutral read on quality across the active models, Hugging Face's Open TTS Leaderboard scores open-source and multilingual TTS, including voice cloning.

Free text-to-speech APIs and free tiers

If you don't want to host a model, these are your free options. They have very different strings attached.

Browser and OS voices. The Web Speech API gives web apps a SpeechSynthesis interface that reads text aloud through the device's built-in synthesizer. It costs nothing, and it's fine for accessibility and read-aloud features. You don't choose the voice, quality varies by browser and operating system, and it runs on the user's device, not your server.

Commercial free tiers. The big TTS APIs all offer some way to start free. The allowance types differ, and the amounts change, so check each pricing page before you plan around one. These were checked on Nov 12, 2026:

  • Google Cloud Text-to-Speech: a monthly free character allowance that varies by voice family. Billing must be enabled, and some newer voice types have no free allowance.
  • Amazon Polly: a Polly Free Tier, then per-character pricing by voice type (Standard, Neural, Long-Form, Generative).
  • Azure AI Speech: a Free (F0) tier with a monthly allowance of neural TTS characters.
  • ElevenLabs: a Free / Pay as you go plan alongside its paid plans, plus a separate startup grants program.
  • Cartesia: a Free plan with monthly credits. Its pricing page lists a commercial use license from the Pro plan up, so read the plan terms before you ship anything built on the free plan.

AssemblyAI. AssemblyAI-TTS-1 is billed at $15 per 1M characters. The pricing page doesn't list a free TTS character allowance today. The TTS docs set the free plan's limit at 5 new streaming sessions per minute, so you can build and test on a free account.

Unofficial wrappers. Libraries like edge-tts and gTTS call consumer text-to-speech endpoints that weren't published as developer APIs, so they can break or be blocked without notice, and using them in a product may conflict with the provider's terms of service.

If you're weighing paid APIs rather than free ones, our comparison of the top text-to-speech APIs and our guide to the best TTS APIs for voice agents cover the field.

What open-source TTS costs in production

Open-source TTS is free to download. It isn't free to run. For speech-to-text, the trade is usually framed as managing a bill versus managing infrastructure. TTS has the same trade, plus a few problems that only show up once a voice is talking to real people.

  • GPU serving. Most of the models above that sound good enough for customers, such as Qwen3-TTS, MOSS-TTS, Orpheus and Dia, want a GPU. You're now running inference servers, autoscaling, cold starts and capacity planning for peaks.
  • Streaming input from an LLM. A voice agent's LLM produces text a token at a time. To keep latency down, TTS has to start speaking before the sentence is finished. Few open-source models document streaming text input, and those that do (VibeVoice-Realtime, Orpheus with input streaming) still leave the buffering and sentence-boundary logic to you.
  • Barge-in and cancel. When a caller interrupts, synthesis has to stop at once and the queued audio has to be dropped. That's session-level plumbing you build around the model.
  • Phone-line audio. Telephony runs on 8 kHz μ-law or A-law. Check each model's output format: if it only produces wideband PCM, you resample and transcode on every call.
  • Reading the hard parts. Order numbers, dates, prices, addresses and confirmation codes are where TTS models slip. A wrong digit in a confirmation code costs more than a slightly flat sentence. Test your own entities, not demo sentences. Our guide on how to evaluate TTS for voice agents shows how.
  • Bursty traffic. Call volume spikes. Self-hosted capacity has to be provisioned for the peak, and you pay for the idle GPUs in between.
  • Model churn and maintenance. The table above is the warning. The original Coqui TTS repo went quiet in 2024, and Chatterbox, a common replacement, was retired in 2026. Each switch is a re-evaluation, a re-integration and sometimes a license review.
  • Latency you can measure. Time to first audio decides whether a conversation feels natural. Our TTS latency guide breaks down where the milliseconds go.

None of this means self-hosting is wrong. For offline devices, strict data-locality rules, research or fine-tuning a unique voice, open-source TTS is the right tool. Just price in the engineering time.

Read the TTS docs

See how a streaming TTS session handles LLM text, flushes, cancel on barge-in and μ-law or A-law output for phone lines, with runnable Python and JavaScript samples.

View the docs

Where a managed TTS API fits

A managed API moves the production list above onto someone else's roadmap. Here's how AssemblyAI-TTS-1 (assemblyai-tts-1) handles each item, using only what's documented:

  • Built for conversation. It's trained on millions of hours of audio, including real conversational data, and it's built for voice agents and live calls rather than narration.
  • Six languages: English, Spanish, French, German, Italian and Portuguese. Thousands of voices are in the catalog. The default English voice is aai-kyla.
  • Streaming in, streaming out. One WebSocket session per call accepts text as your LLM produces it, so you can send partial sentences and flush at sentence ends. A Cancel message drops queued audio on barge-in and keeps the session open. There's also an HTTP API for one-shot requests.
  • Telephony-ready audio. The WebSocket API outputs pcm_mulaw or pcm_alaw (8-bit G.711) directly, alongside 16-bit PCM.
  • Voice cloning with consent. You can clone a voice from 10 to 30 seconds of speech in the dashboard, with the speaker's consent on record. Cloned voices are billed at the same rate. For background, see what voice cloning is.
  • Session limits. Limits count new sessions per minute, not sessions open at once. The starting limit is 5 per minute on a free plan and 200 on a paid plan, and the paid limit grows with use.
  • Pricing: $15 per 1M characters, billed on the input characters you send.

On the benchmarks, AssemblyAI's own evals in the AssemblyAI-TTS-1 launch post, dated Oct 28, 2026, measured 99.0% word accuracy and 93.7% entity accuracy. Entity accuracy covers numbers, codes and abbreviations read correctly. In a six-model head-to-head over each provider's API, AssemblyAI-TTS-1 returned its first byte of audio in 47 ms at the median, and the post quotes about 90 ms time to first audio in production. On Cekura's independent text-to-speech benchmark, as of Nov 10, 2026, it ranked #1 on reading codes and addresses.

For the listening side of the same voice agent, pair it with AssemblyAI-STT-3.7-Realtime-Agent (assemblyai-stt-3-7-realtime-agent), so speech in and speech out come from one vendor.

Want to see it in code first? Our Python text-to-speech tutorial walks through a working build.

How to choose

Start with the license, then maintenance, then sound quality. A model you can't ship, or one nobody maintains, isn't a bargain no matter how it sounds.

If you need… Start with Watch out for
A small, permissive model for an app or prototype Kokoro-82M No cloning, and the repo has been quiet since Aug 2025
Offline speech on a Raspberry Pi or edge device Piper (piper1-gpl) GPL-3.0-or-later engine; check each voice's dataset license
Voice cloning you can use commercially Qwen3-TTS or MOSS-TTS GPU serving; your own consent process for cloned voices
Multi-speaker dialogue for content Dia English only; forbidden-use terms
Research or non-commercial cloning experiments F5-TTS or XTTS-v2 Non-commercial weights; can't move to production
Read-aloud in a web page at zero cost Web Speech API Voice and quality vary by device
A production voice agent on phone lines A managed streaming TTS API such as AssemblyAI-TTS-1 Language coverage: check that your languages are supported

The open-source TTS landscape will look different again in a year. Pick for the license you can live with and a maintainer you can count on, and keep your evaluation set handy for the next switch.

Build your voice agent's voice

Stream LLM text in and get natural, accurate speech out in six languages, with cancel on barge-in and telephony audio built in. Create a free account to start building.

Sign up free

Frequently Asked Questions (FAQs)

What is the best open-source text-to-speech model?

There's no single best open-source TTS model. It depends on your license needs and hardware. Kokoro-82M is the best small, permissive option (Apache-2.0 code and weights), Qwen3-TTS is the strongest permissive choice for voice cloning in 10 languages, and Piper leads for offline and edge devices. Check the weights license and the last commit date before you pick on sound quality alone.

Is there a free text-to-speech API?

Yes. The browser's Web Speech API is free but uses whatever voices the device has. Google Cloud, Amazon Polly, Azure, ElevenLabs and Cartesia all offer free tiers or allowances, with different limits and terms. AssemblyAI's pricing page lists no free TTS character allowance today, but its free plan allows up to 5 new TTS sessions per minute for building and testing.

Can I use open-source TTS models commercially?

Only if the weights license allows it, and the code license doesn't answer that. Kokoro, Qwen3-TTS, MOSS-TTS, Dia and Chatterbox have weights licensed for commercial use. F5-TTS (CC-BY-NC-4.0) and XTTS-v2 (Coqui Public Model License) don't. Fish Audio, IndexTTS-2, Higgs Audio and NeuTTS Nano have custom licenses with revenue thresholds or other limits.

What happened to Coqui TTS?

The original coqui-ai/TTS repository hasn't had a commit since February 2024, and coqui.ai no longer loads. The Idiap Research Institute maintains an active fork, idiap/coqui-ai-TTS, published on PyPI as coqui-tts under MPL-2.0. The fork updates the code, but XTTS-v2's weights are still under the non-commercial Coqui Public Model License.

Is Piper TTS still maintained?

Yes, but not in the original repository. rhasspy/piper is archived, and development moved to OHF-Voice/piper1-gpl, which had a commit on Nov 11, 2026. The new codebase is GPL-3.0-or-later instead of MIT, and each voice's usage rights depend on its training data, so review both before you embed Piper in a product.

Which open-source TTS is best for real-time voice agents?

Qwen3-TTS, MOSS-TTS-Realtime, Orpheus and VibeVoice-Realtime all document streaming output, and VibeVoice-Realtime and Orpheus also document streaming text input. You'll still have to build barge-in handling, telephony audio conversion and GPU autoscaling yourself. If you'd rather not, a managed streaming TTS API like AssemblyAI-TTS-1 handles those at the session level.