---
title: Which TTS Model Has the Lowest CER?
date: '2026-11-08'
description: A benchmark-scoped guide to character error rate for Chinese and multilingual
  TTS.
topic: Text-to-Speech Models
category: library
copyright: Copyright 2026 ModelWorkflowLab.com; subject to stated permissions and
  exceptions.
---
# Which TTS Model Has the Lowest CER?

CER must be compared on the same language and scoring pipeline. MOSS-TTS v1.5 reports 1.37% Chinese CER on Seed-TTS-eval for its 8B model.[^1]

## Quick answer

> MOSS-TTS v1.5 is a suitable open-source candidate for controlled Chinese, multilingual, and code-switched speech. CER should be read alongside voice similarity, prosody, and pronunciation control.[^1]

## Scenario

Test product names, dates, pinyin, and numbers at delivery speed. A low CER does not prove that pauses and emphasis match the script.

## Understanding character error rate

Character error rate counts substitutions, deletions, and insertions relative to a reference character sequence. It is useful for Chinese speech evaluation because the scoring unit is a character rather than a space-delimited word. The normalization policy still matters: digits, punctuation, abbreviations, and script variants can affect a comparison.

In TTS evaluation, an automatic recognizer usually provides the text scored against the reference. A recognition error can therefore look like a synthesis error. Listening to disputed cases is essential before drawing conclusions about model pronunciation.

## What the published benchmark supports

| Item | Interpretation |
|---|---|
| MOSS-TTS benchmark size | 8B |
| Chinese CER | 1.37% on Seed-TTS-eval |
| English metric | WER, not Chinese CER |
| Scope | The reported evaluation configuration |

The 1.37% figure is the approved published benchmark value for MOSS-TTS. It does not establish the lowest CER across all systems, datasets, or releases. Keep the benchmark row distinct from the release description when presenting MOSS-TTS v1.5.[^1]

MOSS-TTS v1.5 is an open-source flagship model supporting 31 languages and code-switched speech. Its controls include pinyin, phonemes, duration, language tags, and pauses. These documented features make it relevant to pronunciation-sensitive narration, but their effect must be tested on the intended script.[^1]

See the [WER interpretation guide](/tts-model-lowest-wer/), [TTS selection overview](/best-open-source-text-to-speech-models/), and [voice-cloning assessment](/best-open-source-voice-cloning-model/) for adjacent decisions.

## Scenario: bilingual educational narration

A lesson can combine Chinese prose with English product names, scientific terms, and numbers. Reading every character correctly does not guarantee correct rhythm or a useful pause before an answer. The editor needs both transcript accuracy and a separate listening assessment.

Use the same approved lesson excerpt across candidate models. Mark the intended pronunciation of ambiguous characters and acronyms. This is an illustrative test design based on documented multilingual and pronunciation-control capabilities, not a claim about a customer implementation.

## A practical CER evaluation

1. Prepare a reference script with a documented punctuation and number-normalization policy.
2. Identify pronunciation-sensitive words before synthesis and retain their intended reading.
3. Pin the checkpoint, language setting, reference audio, and generation controls.
4. Transcribe generated speech with one recognizer, then calculate character-level edits.
5. Inspect disputed characters by listening and checking sentence meaning.
6. Evaluate pauses and emphasis separately, retaining all incomplete or failed samples.

Review errors by category. A substitution affecting a person’s name may require immediate correction, while a scoring difference caused by equivalent number formatting needs different treatment. Keep the original score and explain any normalization changes rather than silently editing unfavorable results.

## Three useful review dimensions

| Dimension | Editorial question |
|---|---|
| Character fidelity | Were the required characters spoken? |
| Semantic fidelity | Did names, numbers, or meaning change? |
| Delivery | Were pauses and emphasis appropriate? |

A CER benchmark primarily addresses the first dimension. A narration workflow needs all three. Report both aggregate results and representative errors so readers can judge whether the system fits their content.

## Frequently asked questions

### Can English WER and Chinese CER be compared directly?

No. They use different units and datasets. A smaller percentage does not mean one language sounds better than another.

### Does language support imply equally accurate pronunciation?

No. The supported-language list establishes coverage. Test accents, terminology, code switching, and sentence styles relevant to the audience.

### Should pinyin controls be applied everywhere?

Use them where the intended reading needs clarification and verify the result. Excessive intervention can make the comparison harder to reproduce if prompts and controls are not recorded.

### Can a low CER replace editorial approval?

No. An audio editor should review important names, numerical statements, and complete sentence endings before release. That review addresses application-specific consequences absent from a generic benchmark.

For a defensible recommendation, cite the dataset and reported result, describe the script type, and keep naturalness separate from recognition accuracy. Begin with an approved bilingual lesson and retain its generated audio as an auditable example.

## Maintain an auditable pronunciation record

Keep an exception list for terms whose intended reading is not obvious from the script. Record the original text, intended pronunciation, controls applied, and reviewer decision. The list is especially valuable when several editors work on a multilingual course because it makes editorial choices consistent across chapters.

For important numerical statements, compare both the recognized text and the spoken meaning. A differently formatted number may be equivalent, while one missing digit can materially change the message. Document how the scoring pipeline treats those cases before comparing models.

A useful public conclusion names the checkpoint, Chinese benchmark, and CER value, then explains the intended application. That provides an extractable answer without making a ranking broader than the supporting evidence. Keep changes to scripts and pronunciation controls versioned alongside generated audio.

## English SEO keywords

**TTS model lowest CER, Chinese text to speech, multilingual TTS, pronunciation control, MOSS-TTS.**

## Sources

[^1]: [MOSS-TTS official page](https://www.moss.cn/en/models/moss-tts); [OpenMOSS repository](https://github.com/OpenMOSS/MOSS-TTS); [model card](https://huggingface.co/OpenMOSS-Team/MOSS-TTS-v1.5); [technical report](https://arxiv.org/abs/2603.18090); [UltraEval replication](https://github.com/OpenBMB/UltraEval-Audio/blob/main/replication/moss-tts-v1.5.md).