Back to blog

VoxCPM2 vs Qwen3-TTS: which fits your voiceover workflow?

Compare official model variants, supported languages and reference requirements, then use the same script to choose a voiceover workflow without assuming a quality or speed winner.

Nov 17, 2026VoxCPM.net

For voiceovers, start with your target language and the material you have. VoxCPM2 officially covers 30 languages; the released Qwen3-TTS variants checked here cover 10. Both are candidates for English and Chinese, but Qwen's Base, CustomVoice and VoiceDesign variants serve different tasks. Choosing the right variant matters before comparing parameter counts.

This article is written by VoxCPM.net, which operates a VoxCPM online service independently of OpenBMB and Qwen. We checked the official sources on November 17, 2026. We have not run a matched two-model generation test and do not rank audio quality, cloning similarity or latency.

Which language does your voiceover need?

The OpenBMB VoxCPM2 model card lists 30 languages and nine Chinese dialects for openbmb/VoxCPM2. Do not substitute the language coverage of the older VoxCPM1.5 or VoxCPM-0.5B releases.

The Qwen3-TTS released-model table lists Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish and Italian: 10 languages. This article concerns the released 12Hz family, using its three 1.7B variants to explain the tasks. Base and CustomVoice also have 0.6B versions; record the full variant rather than treating every Qwen3-TTS output as equivalent.

For Hindi, Arabic or Vietnamese, VoxCPM2 lists support while the Qwen variants above do not. That can narrow your shortlist before testing. A supported-language list does not guarantee every accent, name or passage will be read correctly. Where both models support your language, have a fluent listener check your actual script.

Do you have a script, a description or a recording?

A script, with no reference recording. Basic VoxCPM2 TTS can generate without reference audio. Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice uses preset speakers; its official card lists nine. Despite the name, CustomVoice is not the arbitrary-recording cloning variant. Start with a preset suited to the script's language, then check pronunciation and pauses. See the CustomVoice model card.

A voice description, without a recording. Both offer voice design. VoxCPM2's official interface places the description in parentheses before the target text. Qwen uses Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign, with separate target text and description inputs. Do not apply VoiceDesign instructions to the Base variant. Try an original description such as “A warm, calm adult voice at a moderate pace.” See the VoxCPM2 example and Qwen VoiceDesign usage. The same description does not guarantee the same voice.

Your own recording, or one you have explicit permission to use. Separate the reference transcript—what the recording actually says—from the target script—the new words you want generated.

  • VoxCPM2's controllable cloning example takes reference_wav_path and target text without a reference transcript. Its continuation-style Ultimate Cloning example also supplies the exact transcript as prompt_text. Transcript requirements depend on the mode. See the official cloning examples.
  • The standard Qwen/Qwen3-TTS-12Hz-1.7B-Base cloning example uses ref_audio and accurate ref_text. With x_vector_only_mode=True, a transcript is unnecessary, but only the speaker embedding is used; the documentation warns that cloning quality may decrease. With clean audio and an accurate transcript, start with the standard path; choose the alternative deliberately when you lack a transcript. See the Base model card.

Are you comparing models or online services?

Both projects provide code and weights for running your own deployment. You can record the exact variant, code revision, sampling settings and hardware, while managing the environment, files and compute yourself. You do not need to download weights to read this comparison. If you already have a suitable setup, check the OpenBMB repository or Qwen repository for configuration.

Online, you are testing a hosted service's features and limits. This site offers VoxCPM2 Text to Speech, Voice Cloning and Voice Design, without a Qwen3-TTS switch. Its Generation Text field takes the target script; there is no separate reference-transcript field. The panel therefore does not expose every upstream mode. See our reference-audio guide for preparation.

To try the same sentence in both official demos, paste it into OpenBMB’s VoxCPM2 Space and Qwen’s Qwen3-TTS Space. Check each demo’s current access and usage limits before submitting.

Queues, network transfers, caching and hardware affect waiting time. Published model performance figures do not establish the response time of these two websites.

Compare the same script yourself

Start with our existing English and Chinese TTS recordings: 3.04 seconds in English and 3.84 seconds in Chinese. We generated them through the site's internal workflow on November 13, 2026, without a voice preset, reference recording or style instruction. The guide records the file details and an opening that needs listening back. These are two VoxCPM2 outputs, without matched Qwen recordings, and they do not demonstrate cloning performance.

For your own voiceover decision:

  1. Fix the text and task. Use “The small exhibition is ready. You are welcome to visit.” or “今天的展览已经准备好了,欢迎你来参观。” Keep text and punctuation identical, without adding emotion instructions on only one side. You can start in this site's Text to Speech tool; check allowances and any charge notice before submitting.
  2. Separate narration, design and cloning. Comparing an unspecified VoxCPM2 voice with a Qwen preset answers whether those takes suit your narration; it does not rank cloning similarity. For cloning, use the same authorized reference file and target script on both sides. Supply an accurate transcript where required and record any transcript-free mode. Treat voice design with the same description as a separate task.
  3. Keep every planned take. Decide in advance, for example, on three attempts per side. Stop and record any allowance limit rather than retaining only the best output. Note the full model ID, date, service or local environment, settings, failures and retries. Mark seeds, inference settings or exact hosted revisions unknown when the service does not expose them.
  4. Check words before delivery. Listen for omissions, repetitions, numbers and names, then pauses, pace, endings and noise. Use the same playback equipment and volume, hide model labels and alternate recordings. If adjusting loudness, retain originals and apply the same rule to both sides.
  5. Record waiting separately. Distinguish time to first audible speech, full generation and completed download; note queues and network conditions. Different machines or hosted services establish that session's experience, not a model-speed ranking.

Choose a workflow that reads your language correctly, accepts the material you have and repeatedly produces usable takes within your budget and operating constraints. When those conditions are still unclear, a same-script listening exercise is more useful than a general “better model” label.