XTTS v2 Review (2026): Multilingual Voice Cloning from 6 Seconds of Audio

What XTTS v2 Actually Does

• Few-second voice cloning. A single 6-second reference audio clip is enough to clone a voice, per Coqui's official model card — no fine-tuning or training run is required for a new voice.
• 17-language support with cross-lingual cloning. Supported languages are English, Spanish, French, German, Italian, Portuguese, Polish, Turkish, Russian, Dutch, Czech, Arabic, Chinese (zh-cn), Japanese, Hungarian, Korean, and Hindi — you can clone a voice from English audio and generate speech in any of the other 16 languages with that same cloned voice.
• Streaming inference. XTTS v2 supports streaming synthesis with sub-200ms latency to first audio, documented since the feature was introduced in Coqui TTS v0.20.0 — useful for interactive voice applications where waiting for a full audio file to render is too slow.
• Runs via the Coqui TTS toolkit. The primary supported way to run XTTS v2 is through Coqui TTS (pip install coqui-tts), which exposes it as TTS("tts_models/multilingual/multi-dataset/xtts_v2").
• Speaker embedding reuse. Beyond passing a raw reference clip each time, the toolkit supports computing and reusing a speaker's latent embedding, avoiding recomputation for repeated synthesis with the same cloned voice.

Real Usage Examples

• Reference audio quality matters. A clean 6-second clip of a single speaker with no background noise or music produces a noticeably better clone than a short, noisy, or multi-speaker clip.
• Non-interactive CPML acceptance is required the first time XTTS v2 loads in an unattended environment (Docker, CI) — set COQUI_TOS_AGREED=1 before that first load.

# Install the toolkit
pip install coqui-tts

# Accept the CPML non-interactively (required for Docker/CI; otherwise an
# interactive prompt appears on first load)
export COQUI_TOS_AGREED=1

# Python API: clone a voice and generate speech
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)

# Same-language cloning
tts.tts_to_file(
    text="This is a cloned voice speaking a new sentence.",
    speaker_wav="reference_voice.wav",
    language="en",
    file_path="output_en.wav",
)

# Cross-lingual cloning: clone from English audio, speak Spanish
tts.tts_to_file(
    text="Esta es la misma voz clonada, ahora hablando en español.",
    speaker_wav="reference_voice.wav",
    language="es",
    file_path="output_es.wav",
)

# Streaming synthesis (low latency to first audio)
for chunk in tts.tts_stream(
    text="Streaming audio, chunk by chunk, for interactive use.",
    speaker_wav="reference_voice.wav",
    language="en",
):
    play_audio(chunk)

License and Commercial Use

What XTTS v2 Is Not Good For

• Any commercial product without a confirmed license. The CPML is non-commercial, and there is currently no confirmed, active pathway to a commercial license following Coqui AI's 2023 shutdown. Do not ship XTTS v2 in a paid product, ad-supported app, or client deliverable without independently confirming licensing terms first.
• CPU-only, resource-constrained real-time use. XTTS v2 does support low-latency streaming, but that performance assumes GPU acceleration; on CPU alone it is meaningfully slower than a lightweight engine like Piper, and may not be practical for real-time use on modest hardware such as a Raspberry Pi.
• Cloning a voice from very short or noisy audio. While 6 seconds is the documented minimum, a noisy, compressed, or multi-speaker reference clip produces a noticeably worse clone than a clean, single-speaker recording.
• A language outside its supported 17. XTTS v2 supports exactly English, Spanish, French, German, Italian, Portuguese, Polish, Turkish, Russian, Dutch, Czech, Arabic, Chinese (zh-cn), Japanese, Hungarian, Korean, and Hindi — there is no official roadmap for additional languages as of publication.
• Impersonating someone without consent. Cloning a real person's voice without their knowledge or consent raises consent, right-of-publicity, and potentially fraud or impersonation concerns that are independent of the CPML's non-commercial restriction — those apply regardless of license terms, in personal and commercial use alike.

Alternatives to XTTS v2

Frequently Asked Questions

Verdict

Sources

• XTTS v2 on Hugging Face — the model card: languages, cloning requirements, and license reference.
• XTTS v2 LICENSE.txt — the full Coqui Public Model License (CPML) 1.0.0 text.
• idiap/coqui-ai-TTS on GitHub — the actively maintained toolkit that runs XTTS v2, with usage documentation.
• Coqui TTS Review — PromptQuorum's dedicated review of the toolkit, including its post-shutdown maintenance history.
• Local TTS & Voice Cloning Licenses — full licensing comparison across local TTS engines.

Related Reading

• Coqui TTS Review — the toolkit that runs XTTS v2, including its maintenance status after Coqui AI shut down.
• Local TTS & Voice Cloning Licenses: Piper, XTTS v2, F5-TTS and Coqui — the licensing deep dive across all major local TTS engines.
• ElevenLabs vs Piper vs XTTS v2 — the cloud-vs-local comparison for readers deciding whether to self-host.
• Piper TTS Review — the fast, permissively licensed, CPU-only local TTS alternative.
• Bark TTS Review — the expressive, non-speech-audio local TTS model, reviewed in the same depth.