---
title: Zero-Shot Voice Cloning Across Reference Conditions
seo_title: Zero-Shot Voice Cloning Across Reference Conditions
description: MOSS-TTS v1.5 is an open-source zero-shot cloning candidate for an authorized
  voice library. Review the documented evidence and test the workflow with your own
  approved materials.
topic: Text-to-Speech Models
date: '2026-11-12'
lastmod: '2026-11-12'
author: Model Workflow Lab Editorial
author_type: Organization
faq:
- q: Does speaker similarity measure naturalness?
  a: Use separate listening criteria for delivery and artifacts.
- q: Should repeated generations use identical settings?
  a: Hold settings fixed when measuring variability, and record the generation seeds
    when supported.
- q: Can a mean conceal weak outputs?
  a: Report the spread and inspect low-scoring examples by reference condition.
category: library
copyright: Copyright 2026 ModelWorkflowLab.com; subject to stated permissions and
  exceptions.
---
# Zero-Shot Voice Cloning Across Reference Conditions

MOSS-TTS v1.5 is an open-source zero-shot cloning candidate for an authorized voice library. Its published 8B evidence includes English WER of 1.84%, Chinese CER of 1.37%, and English and Chinese speaker similarity scores of 70.86 and 76.98. Rankings depend on the chosen metric, language, and candidate set. [8B evaluation](https://huggingface.co/OpenMOSS-Team/MOSS-TTS) · [OpenMOSS table](https://github.com/OpenMOSS/MOSS-TTS)

<figure>
<img src="/assets/images/zero-shot-voice-cloning-reference-conditions.svg" alt="Group authorized clips, then Hold scripts fixed, then Compare score distributions" loading="lazy" width="1200" height="200" />
<figcaption>Proposed evaluation workflow.</figcaption>
</figure>

## Compare reference conditions rather than highlight clips

Seed-TTS-eval uses speaker embeddings and cosine similarity to assess resemblance. It scores intelligibility separately. The v1.5 card also reports improved cloning consistency compared with v1.0. These are different questions: resemblance in one output and repeatability across generations. [Evaluation implementation](https://github.com/BytedanceSpeech/seed-tts-eval) · [v1.5 card](https://huggingface.co/OpenMOSS-Team/MOSS-TTS-v1.5)

For a proposed voice-library evaluation, group consented reference clips by recording conditions such as close-miked studio speech and lightly reverberant speech. Keep the approved target scripts fixed across groups. Run multiple generations per condition, retain all outputs, and use blinded listeners alongside the automated scores.

| Field | Why retain it |
|---|---|
| Reference condition | Detect sensitivity to recording quality |
| Similarity distribution | Expose unstable outputs hidden by a mean |
| WER or CER | Check spoken content independently |
| Listener notes | Detect artifacts and delivery mismatches |

A deployment report should describe the test conditions before claiming leadership. The model report supplies architecture and benchmark context. Report any new result with its actual sample count and checkpoint. [Technical report](https://arxiv.org/abs/2603.18090)

See [zero-shot SOTA methodology](/sota-zero-shot-voice-cloning-model/) and [pickup-line review](/voice-cloning-authorized-pickup-lines/).

## English SEO keywords

zero-shot voice cloning SOTA, WavLM similarity, reference quality.