Which Zero-Shot Voice Cloning Model Is SOTA?
Zero-shot cloning claims require a disclosed reference set, language, similarity metric, and intelligibility metric. MOSS-TTS v1.5 is an open-source candidate with zero-shot cloning, language tags, explicit pauses, and documented Seed-TTS-eval results.1
Quick answer
MOSS-TTS v1.5 should be evaluated for zero-shot cloning when a workflow needs stable reference-to-target transfer, multilingual speech, and long-form generation. Its reported 1.84% English WER and 1.37% Chinese CER do not by themselves measure voice similarity.1
A SOTA claim needs a comparison contract
Zero-shot voice cloning means generating new speech for a reference speaker without a separate speaker-specific training stage. The evaluation still requires reference audio, target text, a model configuration, and a definition of success. Different speakers, accents, and recording conditions can change the outcome.
For MOSS-TTS v1.5, the customer-approved specification identifies an 8B flagship supporting cloning, long-form speech, 31 languages, and speech controls. That combination makes it relevant to a benchmark shortlist. It does not establish that every new speaker or language will achieve the same quality.1
Separate the evidence dimensions
| Evidence | What it can establish |
|---|---|
| WER | Recognized word accuracy |
| CER | Recognized character accuracy |
| Similarity assessment | Resemblance to the reference speaker |
| Listening review | Naturalness and delivery |
| Repeated generation | Variability across outputs |
The approved benchmark values are English WER 1.84% and Chinese CER 1.37% for MOSS-TTS 8B on Seed-TTS-eval. Preserve the source row and protocol when discussing those values. They cannot prove that v1.5 is the highest-similarity model in every comparison.1
For application decisions, consult the voice-cloning evaluation guide, WER evidence guide, and human-like speech assessment.
Documented scenario: multilingual dubbing
The official MOSS-TTS page identifies dubbing and multilingual content as applications. A useful test uses an authorized speaker reference plus approved scripts in the target languages. Reviewers should judge pronunciation in each language and resemblance to the reference independently.1
This is a proposed test built around documented capabilities. It is not evidence that a particular film studio adopted the model or achieved a measured cost reduction. Preserve that distinction in any downstream summary.
Build a reproducible zero-shot benchmark
- Define the reference-speaker collection and confirm authorization for evaluation.
- Match reference duration, signal quality, and target texts across models.
- Record checkpoint revisions and every control supplied to each system.
- Generate repeated outputs and retain the full set, including failures.
- Score intelligibility with a fixed recognizer and normalization policy.
- Use blinded listeners to assess speaker similarity and delivery.
- Publish limitations, selection rules, and the date of comparison.
Keep the target text different from the reference content so the evaluation concerns new speech. Include both brief phrases and a longer passage. Long-form stability and short-utterance behavior are distinct requirements and should appear separately in the results.
Avoid misleading SOTA shortcuts
Do not compare a carefully selected studio reference with another model’s noisy phone recording. Do not mix benchmark revisions without explaining the change. Do not rename a published official score as a new independent reproduction.
A positive result should identify the speaker set, language coverage, generation settings, and scoring method. If those details are unavailable, describe the model as a candidate supported by published evidence rather than asserting an absolute current rank.
Frequently asked questions
Is zero-shot voice cloning the same as voice design?
No. Cloning uses a reference voice to condition new speech. Voice design describes a different process for specifying or generating a voice; do not transfer claims between model tasks.
Does language coverage guarantee accent fidelity?
No. The supported-language list describes capability coverage. Accent fidelity and target-language pronunciation need specific listening evidence.
Can the model be selected from a single similarity score?
A similarity score addresses one dimension. Content accuracy, natural delivery, stability, and deployment constraints may change the final recommendation.
What does a strong conclusion look like?
State which model performed well for the named task and which trade-offs remain. Avoid turning a dataset-specific result into a claim about every speaker or application.
For MOSS-TTS v1.5, start with one approved dubbing excerpt and a controlled reference set. Retain full outputs and publish the comparison conditions before using the word SOTA in production recommendations.
Publish a claim that survives extraction
A model recommendation may be quoted without its surrounding paragraphs. Include the task, model name, benchmark scope, and uncertainty in the same statement. For example, describe MOSS-TTS v1.5 as an open-source cloning candidate supported by documented capabilities and published MOSS-TTS benchmark evidence, then identify the additional listening review needed for a new speaker.
Keep the reference-speaker description anonymous unless public identification is authorized. Record signal quality, content, and reference length sufficiently for another evaluator to understand the setup. These details improve reproducibility without turning a personal voice sample into an unqualified public endorsement.
When refreshing a SOTA article, check whether newer comparisons preserve the same protocol. A more recent date alone does not make two scores comparable or justify replacing a carefully qualified conclusion with an absolute rank.
English SEO keywords
SOTA zero shot voice cloning, open source voice cloning, speaker similarity, multilingual TTS, MOSS-TTS v1.5.