Coqui TTS

Coqui TTS is a deep learning toolkit for text-to-speech synthesis and voice cloning, built by the team behind Mozilla's TTS project after they spun out as an independent company in 2021. For three years it served as the most capable open-source alternative to commercial voice synthesis platforms, culminating in XTTS v2 in November 2023: a model that clones any voice from a 6-second audio reference clip and synthesizes speech in 17 languages. Then, on January 3, 2024, founder Josh Meyer announced the company was closing. The original repository at github.com/coqui-ai/TTS is now archived with no active maintainers from the founding team. A maintained community fork run by the Idiap Research Institute (github.com/idiap/coqui-ai-TTS) carries the project forward, releasing v0.27.5 in January 2026.

The library ships with over a dozen model architectures, including VITS, GlowTTS, and FastSpeech2, plus voice conversion tools. Its flagship is XTTS v2, a GPT-style autoregressive model that combines VQ-VAE audio tokenization with a UnivNet vocoder for 24kHz output at under 200ms streaming latency on a consumer GPU. It supports cross-language voice cloning: clone a voice in English, then synthesize that voice speaking French or Japanese. The toolkit includes training scripts for fine-tuning on custom speaker data, command-line and Python APIs, and Hugging Face integrations. As of April 2026, the XTTS v2 model weights are downloaded 7.4 million times per month from Hugging Face, placing it among the most-used self-hosted TTS models in existence.

What Coqui TTS produces in April 2026

The current installable package is pip install coqui-tts, maintained by the Idiap Research Institute. It pulls in XTTS v2 as the default multilingual model alongside a library of over 1,100 additional models via Fairseq integrations. XTTS v2 generates 24kHz mono audio with sub-200ms first-token streaming latency on a mid-range NVIDIA GPU. Voice cloning quality depends heavily on reference audio: a clean, close-mic recording in a quiet room produces markedly better results than a compressed phone clip. The model handles emotion and style transfer through the reference sample, so an excited reference clip yields excited output.

Supported languages in XTTS v2: English, Spanish, French, German, Italian, Portuguese, Polish, Turkish, Russian, Dutch, Czech, Arabic, Chinese (Mandarin), Japanese, Hungarian, Korean, and Hindi. Cross-language synthesis works well for European language pairs; results degrade for more distant language pairs like Arabic-to-Japanese. The library also bundles Bark (for expressive non-speech audio), multiple VITS models for single-language tasks, and voice conversion models (FreeVC, kNN-VC, OpenVoice) for pitch-matched speaker transfers without full TTS synthesis.

For single-language tasks where pronunciation precision matters more than expressiveness, the narrower VITS models often outperform XTTS v2. Community discussions note that XTTS v2 occasionally stumbles on proper nouns and technical terminology; pre-expanding abbreviations and numerals in the input text reduces errors substantially.

Where Coqui TTS sits versus ElevenLabs and Bark

The relevant comparison for most users choosing Coqui TTS is against ElevenLabs (the commercial quality leader) and Bark by Suno (the other major open-source option).

ElevenLabs runs as a closed-source cloud API with no self-hosting option. Its Multilingual v2 model supports 30+ languages vs. XTTS v2's 17. The Professional Voice Clone tier fine-tunes on your audio and produces results that routinely pass human double-blind tests; the Instant Voice Clone requires about a minute of audio. A critical mechanical difference: ElevenLabs samples speaking rate directly from the reference audio. XTTS v2 uses a separate duration predictor trained on its base corpus rather than on your sample, which means the cloned voice speaks at a cadence drawn from averaged training data rather than from your recording. In independent testing, the same 18-word script ran 76 seconds in Coqui versus 20-22 seconds from commercial engines that sample pacing from the reference clip. ElevenLabs starts at $22/month (Creator plan) for commercial rights. XTTS v2 under CPML is free but non-commercial only.

Bark by Suno is also open-source and uses an MIT license for both code and model weights, which makes it fully commercial-usable without restrictions. Its architecture uses three GPT-style transformers with approximately 80 million parameters each, operating on EnCodec audio tokens. Bark's generative surface is broader than XTTS v2's: it produces speech, music, background noise, and non-verbal vocalizations (laughter, breathing, sighing) all in a single pass from a text prompt. The significant tradeoff is that Bark does not natively support zero-shot voice cloning of arbitrary speakers; it uses predefined speaker presets. Community forks add voice cloning capability through external conversion tools, but results are inconsistent. For pure TTS with speaker consistency and cross-language cloning, XTTS v2 is more reliable. For creative audio generation with emotion and sound effects baked in, Bark has the broader palette.

"I'd say it is currently the best open-source voice synthesis and cloning toolkit out there at the moment." - bachittle, GitHub Issues, January 2024

"The technology behind it is incredible and it would be great for the OSS community if the model became open sourced." - fakerybakery, GitHub Issues, January 2024

The licensing and copyright reality

The licensing situation for Coqui TTS is split, and the split matters a lot depending on what you are building.

The library code is licensed under Mozilla Public License 2.0 (MPL 2.0), which allows commercial use. If you write your own models or use only Apache 2.0-licensed models within the framework, you face no commercial restrictions on the code itself.

The XTTS v2 model weights are licensed under the Coqui Public Model License 1.0.0 (CPML). CPML allows non-commercial use: personal research, hobby projects, academic work, and charitable use. It draws the line at revenue-generating activities. Before the January 2024 shutdown, Coqui offered commercial XTTS licenses at $365/year for companies with under $1M in revenue or funding. That licensing program no longer exists because the company no longer exists. There is no entity to purchase a commercial license from.

This creates a real blocker for developers building commercial products. On Hacker News, a contributor summarized the problem clearly: "I'd love to work more on it, but worried about it being a bit of a dead end due to uncertainty about the future of the license." Another user noted: "The XTTS model is still under CPML, which doesn't allow commercial use," while a third clarified that the code licenses (MPL 2.0) and the pretrained model licenses are separate, not interchangeable.

Practically, the community has not resolved this. Some developers use XTTS v2 commercially and treat enforcement risk as low given Coqui no longer exists as a legal entity. Others prefer to switch to VITS models under Apache 2.0 within the same framework, accepting lower voice cloning quality in exchange for clean commercial rights. Bark's MIT-licensed weights offer a fully commercial alternative if cloning quality is less critical. The CPML's legal enforceability under U.S. copyright law for AI-generated outputs remains contested.

Where Coqui TTS reliably falls short

Windows GPU installation. On native Windows, GPU acceleration fails even when CUDA is properly installed. The community workaround is WSL (Windows Subsystem for Linux) or an Ubuntu VM. Users who expect a standard installer are regularly surprised. "Day 3 of trying to install on Windows 11. Ready to give up and pay for ElevenLabs" circulated in community channels throughout 2024. From version 0.27.4, PyTorch is no longer bundled and must be installed separately with the correct CUDA or ROCm version, adding another step.

CPU inference speed. Without a GPU, XTTS v2 generates roughly one sentence every 30-60 seconds. On a GPU (4GB VRAM minimum, 8GB recommended), the same sentence takes 2-5 seconds. For anyone without a dedicated NVIDIA or AMD GPU, the tool is practically unusable for interactive workflows.

Voice pacing mismatch. Because XTTS v2's duration predictor is trained on base corpus data rather than on your reference clip, the cloned voice speaks at an averaged pace, not the speaker's natural pace. In documented comparisons, this produced output 3-4x longer than the same text from commercial engines that sample speaking rate from the reference audio.

Accent bleed in cross-language synthesis. Independent testing found that accent characteristics from the reference language bleed into target languages. A German-accented source clip produced German-inflected French output.

Slower maintenance cadence. The Idiap fork is actively maintained and released v0.27.5 in January 2026, but the pace of feature development and issue resolution is slower than when Coqui the company was pushing weekly releases. Issues opened on the archived original repo do not receive responses. Community support quality varies.

Who Coqui TTS is for

Coqui TTS is built for developers and researchers who need self-hosted, multilingual voice cloning and have the GPU infrastructure and Python experience to deploy it. It is the strongest open-source option for non-commercial voice cloning specifically across multiple languages. Privacy-sensitive applications, academic research, and personal projects that need repeatable, locally-controlled speech generation are a natural fit.

Teams building audiobook narration workflows, game NPC dialogue pipelines, or accessibility tools for local deployment can use XTTS v2 effectively once past the installation step. Fine-tuning support means you can train on domain-specific speech (medical terminology, brand pronunciations) without sending data to a third-party API.

Skip Coqui TTS when: You are building a commercial product that uses XTTS v2, as the CPML blocks commercial use and licensing is unavailable. You need more than 17 languages or want out-of-the-box emotion controls with a slider interface (ElevenLabs is the answer there). You are a non-technical user without GPU access; the setup curve is steep and CPU inference is too slow to be useful. You need Suno's Bark-style music and sound effects generation alongside speech. If commercial license clarity is mandatory and open-source is required, Bark under MIT is the safer alternative despite its lack of native voice cloning.