Coqui TTS Review (2026): The Community-Maintained Voice-Cloning Toolkit

History: From Mozilla TTS to a Community Fork

What Coqui TTS Actually Does

• Multi-model interface. A single TTS() Python class loads and runs different model architectures, including VITS-based single-speaker and multi-speaker models, Tacotron2-based models, and the XTTS v2 voice-cloning model, without changing the surrounding application code.
• Command-line synthesis. The tts CLI command synthesizes speech directly from the terminal, listing all available pretrained models with tts --list_models.
• Voice cloning via XTTS v2. Passing a speaker_wav argument to a loaded XTTS v2 model clones a voice from a short reference audio clip, generating speech in that cloned voice across the languages XTTS v2 supports.
• Broad pretrained model library. The toolkit ships with access to pretrained models described as covering over 1,100 languages via the underlying Fairseq-based massively multilingual models, alongside higher-quality curated models for specific languages.
• Model training and fine-tuning. Beyond running pretrained models, the toolkit includes training scripts and dataset-analysis utilities for building or fine-tuning a custom voice model, historically one of its most-used features before the shutdown of Coqui AI's hosted training service.
• Streaming inference for XTTS v2. The toolkit documents low-latency streaming synthesis with XTTS v2 (cited by the maintained fork's documentation as sub-200ms latency to first audio), useful for interactive voice applications.

Coqui TTS and XTTS v2: How They Relate

Install and Run Coqui TTS: Step by Step

1. 1
   Install the maintained package.
   Why it matters: Run `pip install coqui-tts` (or `uv pip install coqui-tts` if using uv) in a Python environment. Install the `coqui-tts` package specifically — not the older `TTS` package name, which points to the now-unmaintained original repository.
2. 2
   List available pretrained models.
   Why it matters: Run `tts --list_models` to see the full catalog of pretrained models available to download, organized by language and architecture (VITS, Tacotron2, XTTS, and others).
3. 3
   Synthesize speech from the command line.
   Why it matters: Run `tts --text "Hello world" --out_path output.wav` to synthesize with the default model, or add `--model_name <model>` to pick a specific one from the list.
4. 4
   (Optional) Use the Python API for XTTS v2 voice cloning.
   Why it matters: Load the model with `TTS("tts_models/multilingual/multi-dataset/xtts_v2").to(device)`, then call `.tts_to_file()` with a `speaker_wav` argument pointing to a short reference audio clip and a `language` argument to clone a voice.
5. 5
   (Optional) Enable GPU acceleration.
   Why it matters: Pass `gpu=True` to the `TTS()` constructor, or call `.to("cuda")` on the loaded model object, to run inference on an NVIDIA GPU — recommended for XTTS v2, which is significantly slower on CPU alone.
6. 6
   (Optional) Accept the CPML license non-interactively for XTTS v2.
   Why it matters: Loading XTTS v2 for the first time prompts an interactive license-acceptance step for the CPML. Set the `COQUI_TOS_AGREED=1` environment variable to accept it non-interactively, which is required for unattended use in Docker containers or CI pipelines.

Real Usage Examples

• Non-interactive CPML acceptance for automated environments: set COQUI_TOS_AGREED=1 before the first XTTS v2 load, so Docker builds and CI pipelines do not hang on an interactive prompt.
• Speaker and language introspection: after loading a multi-speaker or multilingual model, tts.speakers and tts.languages list what the loaded model actually supports, useful for validating input before synthesis.

# Command line: list available pretrained models
tts --list_models

# Command line: synthesize with the default model
tts --text "Hello world" --out_path output.wav

# Command line: synthesize with a specific model
tts --model_name "tts_models/en/ljspeech/tacotron2-DDC" \
    --text "This is a test." --out_path output.wav

# Python API: basic synthesis
from TTS.api import TTS

tts = TTS("tts_models/en/ljspeech/tacotron2-DDC")
tts.tts_to_file(text="Hello world", file_path="output.wav")

# Python API: XTTS v2 voice cloning (accept CPML non-interactively first)
# export COQUI_TOS_AGREED=1
import torch
from TTS.api import TTS

device = "cuda" if torch.cuda.is_available() else "cpu"
tts = TTS("tts_models/multilingual/multi-dataset/xtts_v2").to(device)

tts.tts_to_file(
    text="Hola, esto es una prueba.",
    speaker_wav="reference_voice.wav",
    language="es",
    file_path="cloned_output.wav",
)

# Python API: list speakers and languages available on a loaded model
print(tts.speakers)
print(tts.languages)

License and Cost

What Coqui TTS Is Not Good For

• Building on the original, unmaintained repository. If you pip install TTS (the original package name) rather than pip install coqui-tts (the maintained fork), you are building on code that has received no updates since August 2024 — check which package a tutorial or dependency actually references before relying on it.
• Fastest possible CPU-only synthesis. If your priority is real-time speech on modest hardware like a Raspberry Pi with no GPU, Piper is purpose-built for that and simpler to deploy; Coqui TTS's larger models, especially XTTS v2, are noticeably slower on CPU alone.
• Commercial voice cloning without a separate license check. Running XTTS v2 through the Coqui TTS toolkit still carries XTTS v2's own non-commercial CPML restriction on the model weights and outputs — the toolkit's MPL-2.0 license does not override that. See the License and Cost section above.
• Guaranteed long-term corporate support. Coqui AI, the company, no longer exists as of December 2023. The current fork is maintained by a research institute on a community basis, which is a different maintenance model than a funded company with a support contract — factor that into decisions for critical production infrastructure.
• A single, stable API across versions. Because development moved between an unmaintained original and an actively developed fork, some tutorials, Stack Overflow answers, and blog posts online reference an older API surface or the original package name — verify against the current coqui-tts documentation rather than an older search result.

Alternatives to Coqui TTS

Frequently Asked Questions

Verdict

Sources

• idiap/coqui-ai-TTS on GitHub — the actively maintained fork: README, docs, license, and release history.
• coqui-tts on PyPI — published package metadata, including the current MPL-2.0 license and version history.
• coqui-ai/TTS on GitHub — the original, now-unmaintained repository (MPL-2.0 license).
• XTTS v2 on Hugging Face — the voice-cloning model's model card and CPML license text.
• Idiap Research Institute — the Swiss research institute maintaining the community fork.

Related Reading

• Local TTS & Voice Cloning Licenses: Piper, XTTS v2, F5-TTS and Coqui — the licensing deep dive across all major local TTS engines.
• XTTS v2 Review — a full breakdown of the voice-cloning model this toolkit can run.
• Piper TTS Review — the fast, CPU-only local TTS alternative, reviewed in the same depth.
• ElevenLabs vs Piper vs XTTS v2 — the cloud-vs-local comparison for readers deciding whether to self-host.
• Bark TTS Review — the expressive, non-speech-audio local TTS model, reviewed in the same depth.