# Ming-3.0 speech evaluation: Voice Benchmark Profile

> inclusionAI reports 0.800% Chinese and 1.584% English word error rate on the Seed-TTS testset for Ming-3.0-tts-0.5B, and 3.45 Chinese CER with 3.69 English WER for Ming-3.0-audio-flash.

- Scope: Zero-shot, multilingual, paralinguistic, dialect, emotion and podcast text-to-speech; Chinese, English, dialect, context and long-form speech recognition; audio understanding, spoken interaction and tool calling
- Measurement lanes: Voice experience, Spoken reasoning
- Primary metric: Word and character error rates (%, lower is better); accuracy, similarity and benchmark scores (higher is better)
- Available evidence: Exact values in the Hugging Face model cards; the audio and ASR cards publish them as table images
- Owner: inclusionAI
- Source snapshots refreshed: 2026-10-19

- Provider report retrieved: 2026-10-21

## Zero-shot text to speech on the Seed-TTS testset

inclusionAI-reported. WER is lower-is-better; SIM (speaker similarity) is higher-is-better.

| Model | Chinese WER | Chinese SIM | English WER | English SIM |
| --- | --- | --- | --- | --- |
| [Ming-3.0-tts-0.5B](/models/ming-3-0-tts-0-5b) | 0.800% | 0.758 | 1.584% | 0.691 |
| [Ming-3.0-tts-tiny](/models/ming-3-0-tts-tiny) | 0.820% | 0.781 | 1.666% | 0.728 |

## Multilingual text to speech

inclusionAI-reported on the MiniMax TTS-Multilingual test set. German, Spanish and French use WER; Korean and Japanese use CER; each is followed by SIM.

| Model | German WER | German SIM | Spanish WER | Spanish SIM | French WER | French SIM | Korean CER | Korean SIM | Japanese CER | Japanese SIM |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| [Ming-3.0-tts-0.5B](/models/ming-3-0-tts-0-5b) | 0.365% | 0.753 | 1.099% | 0.763 | 4.264% | 0.675 | 1.728% | 0.793 | 3.765% | 0.765 |
| [Ming-3.0-tts-tiny](/models/ming-3-0-tts-tiny) | 0.325% | 0.791 | 0.971% | 0.797 | 3.470% | 0.733 | 6.000% | 0.797 | 3.427% | 0.780 |

## Style, dialect, emotion and podcast text to speech

inclusionAI-reported. Dialect CER is lower-is-better; the other columns are higher-is-better except podcast CER. The podcast test covers the tiny model only.

| Model | InstructTTSEval-ZH voice design | Basic attribute control | CV3 emotion control | WSC-Eval-TTS-easy CER | WSC-Eval-TTS-hard CER | Podcast CER |
| --- | --- | --- | --- | --- | --- | --- |
| [Ming-3.0-tts-0.5B](/models/ming-3-0-tts-0-5b) | 72.29% | 93.30% | 72.33% | 2.27% | 3.35% | Not reported |
| [Ming-3.0-tts-tiny](/models/ming-3-0-tts-tiny) | 75.37% | 88.23% | 74.00% | 2.53% | 3.11% | 2.03% |

## Speech recognition

inclusionAI-reported. Chinese and dialect averages are CER; English averages and the context and long-form columns are WER; lower is better. Read from images on the model cards.

| Model | Chinese average CER | English average WER | Dialect average CER | ContextASR WER | Earnings22 WER |
| --- | --- | --- | --- | --- | --- |
| [Ming-3.0-audio-flash](/models/ming-3-0-audio-flash) | 3.45 | 3.69 | 9.40 | 1.52 | 11.60 |
| [Ming-3.0-audio-tiny](/models/ming-3-0-audio-tiny) | 3.79 | 4.36 | 10.29 | 2.12 | 18.29 |
| [Ming-3.0-asr-tiny](/models/ming-3-0-asr-tiny) | 3.55 | 3.99 | Not reported | Not reported | Not reported |

## Audio understanding and spoken interaction

inclusionAI-reported; higher is better. Read from images on the audio model cards. The ASR model does not report these.

| Model | BigBenchAudio | MMAU | MMAR | MMSU | MMAU-Pro | VoiceBench | URO-Bench-Pro | BFCL-v3 Audio | SpeechFCEval |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| [Ming-3.0-audio-flash](/models/ming-3-0-audio-flash) | 99.40 | 82.00 | 74.40 | 78.82 | 66.41 | 86.74 | 81.92 | 81.40 | 82.60 |
| [Ming-3.0-audio-tiny](/models/ming-3-0-audio-tiny) | 99.20 | 75.90 | 67.70 | 73.06 | 56.42 | 74.29 | 54.15 | 79.80 | 69.80 |

## Interpretation limit

All results are inclusionAI-run or developer-reported by inclusionAI. Text-to-speech figures use a unified ASR-based WER or CER protocol over each open model at its recommended settings; speaker similarity (SIM) trails content accuracy, and Korean CER for Ming-3.0-tts-tiny (6.000%) is the clearest multilingual gap. The audio and ASR figures are read from images on the model cards; results marked with an asterisk there are self-evaluated. Competitor rows are not carried here. These results do not enter text-model or voice-agent rankings.

## Primary sources

- [Code](https://github.com/inclusionAI/Ming-3.0-tts)
- [Owner page](https://huggingface.co/inclusionAI/Ming-3.0-tts-tiny)
- [Ming-3.0-tts-0.5B model card](https://huggingface.co/inclusionAI/Ming-3.0-tts-0.5B)
- [Ming-3.0-tts-tiny model card](https://huggingface.co/inclusionAI/Ming-3.0-tts-tiny)
- [Ming-3.0-audio-flash model card](https://huggingface.co/inclusionAI/Ming-3.0-audio-flash)
- [Ming-3.0-audio-tiny model card](https://huggingface.co/inclusionAI/Ming-3.0-audio-tiny)
- [Ming-3.0-asr-tiny model card](https://huggingface.co/inclusionAI/Ming-3.0-asr-tiny)

Canonical page: https://benchlm.ai/voice-benchmarks/ming-3-0-speech-evaluation