---
title: Low-CER Mandarin TTS for Names and Polyphones
seo_title: Low-CER Mandarin TTS for Names and Polyphones
description: MOSS-TTS 8B reports 1.37% Chinese CER and 76.98 Chinese speaker similarity
  in its published Seed-TTS-eval comparison. 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: Can CER catch every pronunciation error?
  a: Use listening review for names and ambiguous readings. CER scores recognized
    characters.
- q: Does MOSS-TTS have the lowest CER in its published table?
  a: The table includes lower CER results. Evaluate identity and control as separate
    requirements.
- q: When should a reviewer specify Pinyin?
  a: Use it for ambiguous readings after confirming the intended pronunciation in
    context.
category: library
copyright: Copyright 2026 ModelWorkflowLab.com; subject to stated permissions and
  exceptions.
---
# Low-CER Mandarin TTS for Names and Polyphones

MOSS-TTS 8B reports **1.37% Chinese CER and 76.98 Chinese speaker similarity** in its published Seed-TTS-eval comparison. The same table reports 0.93% CER for VoxCPM. MOSS-TTS is worth evaluating when the script needs reference-voice similarity and control over specific readings. [OpenMOSS benchmark](https://github.com/OpenMOSS/MOSS-TTS)

<figure>
<img src="/assets/images/low-cer-mandarin-names-polyphones.svg" alt="Approve intended readings, then Generate both variants, then Listen to flagged words" loading="lazy" width="1200" height="200" />
<figcaption>Proposed evaluation workflow.</figcaption>
</figure>

## Verify the sound behind the character

Seed-TTS-eval documents character-level scoring on Chinese text. A recognizer may infer the correct character from context even if a listener hears an unintended pronunciation. The v1.5 model card documents Pinyin controls. Use them to specify readings that a fluent reviewer has already approved. [Seed-TTS-eval](https://github.com/BytedanceSpeech/seed-tts-eval) · [v1.5 controls](https://huggingface.co/OpenMOSS-Team/MOSS-TTS-v1.5)

A proposed Mandarin museum audio guide can include surnames, place names, and polyphones. Mark the expected reading, generate a baseline, and regenerate the flagged words with approved pronunciation input. Have the reviewer listen to the phrase in its full sentence, since natural stress may change when a word is isolated.

| Review field | Evidence to retain |
|---|---|
| Character fidelity | CER and normalization rules |
| Correct reading | Marked script and listening notes |
| Speaker similarity | Reference clip and scoring protocol |
| Phrase flow | Review of the complete sentence |

Report pronunciation fixes separately from recognizer edits. Record both initial and corrected outputs so later editors can see which control was useful. Use the same scoring pipeline for comparisons, and keep the published 8B row distinct from your own v1.5 result. [8B model card](https://huggingface.co/OpenMOSS-Team/MOSS-TTS) · [Official evaluation](https://www.moss.cn/models/moss-tts)

See [CER interpretation](/tts-model-lowest-cer/) and [exact identifier checks](/low-wer-tts-product-codes/).

## English SEO keywords

Mandarin CER, polyphone pronunciation, Pinyin control.