Open-Source, Open-Weight, and Self-Hosted TTS Models: 2026 Guide

Key takeaways

• Downloadable weights are not automatically open source or commercially usable.
• KugelAudio 0 Open is an MIT-licensed checkpoint; the commercial Kugel 3 on-premise service is a separate offer.
• Kokoro-82M and Fun-CosyVoice3-0.5B-2512 publish Apache 2.0 checkpoints. Voxtral-4B-TTS-2603 is CC BY-NC 4.0, and Fish Audio S2 Pro requires a separate licence for any commercial use.
• German is explicitly listed for KugelAudio 0 Open, Voxtral, CosyVoice 3, Fish Audio S2 Pro, Piper and Kugel 3. Kokoro's official pipeline does not list German.
• Published memory figures are runtime-specific. Parameter count is not a safe substitute for measured memory, latency or capacity.
• Streaming support is a serving-stack property as well as a model property. Confirm cancellation, backpressure and time to first playable audio on your target runtime.
• Self-hosting gives the customer control of the data path, but it also transfers patching, capacity, monitoring and incident ownership.

Last updated 10 August 2026. Model cards, repositories and licences were checked on that date.

What is the best self-hostable TTS model in 2026?

There is no defensible universal winner. The useful first decision is whether the deployment needs permissive commercial rights, German, true text-in and audio-out streaming, a small local runtime, or a supported on-premise service. Those gates remove unsuitable options before a listening test.

This table assigns shortlist roles, not quality ranks. No common German test set, voice, hardware or serving configuration was run for this article. For a reproducible evaluation structure, use the open German TTS benchmark.

What is the difference between open source and open weights?

Open-source software provides source code under an open-source licence. Open weights only means that model parameters can be downloaded. The weight licence may still prohibit commercial use, restrict redistribution, require attribution, or impose an acceptable-use policy.

Check four artefacts separately: inference code, model checkpoint, bundled voices or reference audio, and any fine-tuned derivative. For example, Piper's current runtime is GPL-3.0, while its own voice documentation says that each voice's MODEL_CARD contains the relevant licence. Running GPL software inside a company is not the same event as distributing it in a product; distribution and modification can create source-code obligations. Get legal review for the actual packaging rather than labelling GPL as either "commercial" or "non-commercial."

The non-commercial gates are clearer. Voxtral's model card assigns CC BY-NC 4.0 to the checkpoint because of its provided reference voices. The Fish Audio Research License defines internal business use, hosted services and revenue-generating products as commercial and requires a separate written agreement. Neither is an open-source commercial checkpoint.

Apache 2.0 is permissive, but it does not grant rights to a voice actor's identity, customer reference recordings or every third-party dependency. Keep consent and provenance records for every production voice.

Which self-hostable TTS is suitable for German?

Language evidence narrows the list but does not prove production quality. KugelAudio 0 Open, Voxtral, Fun-CosyVoice 3, Fish Audio S2 Pro, Piper and Kugel 3 explicitly list German. Kokoro's official language codes cover American and British English, Spanish, French, Hindi, Italian, Japanese, Brazilian Portuguese and Mandarin Chinese, but not German.

For German, test compounds, abbreviations, dates, currencies, names and code-switching. Test Austrian and Swiss usage separately from standard German, and treat dialect support as a separate requirement. A generic de label says nothing about Bavarian, Swabian or Low German. Our German and dialect selection guide explains how to build that listening set.

What hardware do self-hosted TTS models need?

Do not turn parameter count into a made-up GPU requirement. Memory also depends on precision, vocoder, KV cache, reference-audio length, batch size, CUDA graphs and the serving runtime. Voxtral publishes a clear minimum of 16 GB GPU memory for its BF16 weights with the documented vLLM-Omni path. KugelAudio 0 Open's card documents roughly 19 GB VRAM for its reference implementation and notes that stripping unused encoder weights can reduce that figure.

Fish Audio's current vLLM-Omni recipe is useful as a reference configuration, not a minimum: it records roughly 49 GiB peak use on an A800 80 GB. CosyVoice and Kokoro publish model sizes but no comparable minimum-memory statement. Piper voices range from small local models to larger voice packages, so measure the exact .onnx file and runtime.

For production, capture peak memory and first playable audio at the required concurrency. Then test long inputs, cloned-voice references, cancellation and out-of-memory recovery. Reserve capacity for a warm replica or accept the recovery time after a device failure.

Does documented streaming support guarantee low latency?

No. Voxtral documents streaming and batch inference through vLLM-Omni. CosyVoice 3 documents text-in and audio-out bi-streaming. Fish Audio S2 Pro ships an SGLang streaming path, and the commercial Kugel 3 service documents both input and output streaming. KugelAudio 0 Open's current model card shows batch generation instead. These statements establish a supported mechanism, not a latency guarantee on your hardware.

Kokoro's Python API yields generated segments, and Piper's Python API yields audio chunks, but that alone does not prove incremental text input, backpressure, fast cancellation or stable latency under concurrency. Test the complete transport and audio player. Vendor benchmark numbers are not directly comparable when hardware, input length, reference audio and the definition of latency differ.

What does self-hosted TTS really cost?

Use a one-year workload, not a single GPU-hour price. Include idle replicas, model storage, image scanning, upgrades, monitoring, alerting, load tests, security response and the engineer who owns incidents. Add voice acquisition, pronunciation work and licence review.

Compare alternatives at the same audio volume, peak concurrency, availability target and support level. A permissive checkpoint may have the lowest licence cost and the highest engineering cost. A commercial self-hosted offer may cost more in fees but remove serving work while preserving the customer's network boundary.

How should teams shortlist and deploy a model?

1. Record the exact repository commit, checkpoint ID and voice asset.
2. Approve the code, weight, voice and output terms for the intended use.
3. Reject candidates without the required language or deployment interface.
4. Run one shared pronunciation and native-listener set.
5. Measure first playable audio, throughput, peak memory and error rate on target hardware.
6. Test cancellation, overload, restart, rolling upgrade and rollback.
7. Document logs, traces, caches, backups, telemetry and support access.
8. Assign an owner for security patches, model upgrades and incidents.

Pin every approved artefact by version or digest. A repository moving to a new organisation or licence should not silently change a production build. The current Fish Speech repository now uses S2 Pro in its quick start, so the older S1-mini is not used in this guide. The CosyVoice project now lives under the QwenAudio organisation, although its old FunAudioLLM URL redirects there. Piper is the clearest licence example: the former MIT-licensed rhasspy/piper repository was archived in October 2025, development moved to the GPL-3.0 OHF-Voice/piper1-gpl repository, and that current project is seeking maintainers.

For privacy-sensitive deployments, also audit the full data path. Self-hosting can keep text, audio and reference voices inside customer infrastructure, but logs or external telemetry can defeat that design. See the on-premise infrastructure and data-control guide and the production-readiness checklist.

KugelAudio's documented self-hosted route ships to Kubernetes with a Helm chart, licence key and rollout support. It is a commercial option, not an open-source model. Review the deployment documentation or contact the team.

FAQ

What is the best open-source TTS?

There is no quality winner in this guide. Kokoro is the smallest permissive baseline listed, KugelAudio 0 Open focuses on European languages, and CosyVoice 3 explicitly lists German and bi-streaming. Choose only after a shared listening test and deployment measurement.

Can I self-host text-to-speech commercially?

Yes, when all relevant licences allow the intended use. Apache 2.0 checkpoints such as Kokoro and CosyVoice permit commercial use subject to their terms. Voxtral's published checkpoint is non-commercial, and Fish Audio requires a separate commercial agreement.

Which TTS runs on CPU?

Piper is a local ONNX runtime, and Kokoro's official demo includes a CPU path. That proves execution, not production throughput. Benchmark the exact runtime, voice, text length and concurrency before choosing hardware.

What is the best self-hostable German TTS?

No German quality result is reported here. Start with candidates that officially list German, then test normalization, names, compounds, code-switching and the required regional variety with native listeners.

Does self-hosted TTS provide zero retention?

It enables customer-controlled retention; it does not prove zero retention by itself. Check application logs, traces, caches, backups, crash dumps, telemetry, licence checks and support access.

Is Voxtral TTS commercially open source?

No. The official Voxtral-4B-TTS-2603 model card labels the checkpoint CC BY-NC 4.0. Treat it as a non-commercial open-weight release unless you obtain separate rights that cover the planned use.