Imagine producing an audio guide whose narrator must sound welcoming at the entrance, restrained beside a memorial, and brisk when giving directions. Keeping a recognizable voice while changing its delivery is a different request from imitating one recording. VoxCPM2 gives those requests different inputs.

The project grew from a collaboration between ModelBest and the human–computer speech interaction laboratory at Tsinghua Shenzhen International Graduate School. Tsinghua’s September 2025 account documents the original VoxCPM release.[1] The repository dates VoxCPM2 to April 2026; its technical report followed on June 5, 2026.[2][4] This digest examines that release and the documentation available on September 22, 2026.

The headline specifications are substantial: 2 billion parameters, 30 languages, nine Chinese dialects, and 48 kHz output.[2] The more interesting change is how a user tells the model whose voice to generate and how that voice should speak.

Three ways into the recording booth

The model card distinguishes three practical starting points. With voice design, a written description specifies an imagined voice without supplying a recording. With controllable cloning, a reference clip supplies the voice while an optional instruction requests a delivery style. With continuation cloning, recorded speech and its matching transcript provide the material from which the model continues.[3]

For the hypothetical audio guide, these modes answer different production questions. During casting, a description could explore possible narrators. Once a producer has an authorized reference recording, controllable cloning could test a calmer or faster delivery. Continuation offers another route when the performance in an existing clip is itself the desired model. These are proposed uses of the controls, not listening-test results.

The interface makes the distinction unusually legible. reference_wav_path supplies a recording without requiring its transcript. The older continuation route uses prompt_wav_path together with prompt_text. Voice and style descriptions go in parentheses before the words to be spoken.[6]

That change removes one preparation step from reference-based cloning: accurately transcribing the reference. It also asks a more precise question of the recording. Should the model use it as evidence of speaker identity, or as the beginning of a performance to carry forward? A short clip can serve either purpose, but the intended relationship to the new speech matters.

What changed inside

The technical report explains the new reference pathway as a marked segment at the beginning of the input. Later generation can attend to it, but it need not be the immediately preceding passage of speech. The authors use this arrangement to separate reference-based speaker conditioning from instructions about delivery. Continuation remains available, and the two pathways can be combined.[4]

This extends the original model’s architecture. The September 2025 paper describes a hierarchy in which one component models semantic and prosodic structure while a residual component handles finer acoustic detail. A local diffusion decoder generates continuous speech representations, and the system trains jointly rather than relying on an external discrete speech tokenizer.[5]

“Tokenizer-free” needs that qualification. It does not mean the system dispenses with text tokens or every form of quantization: the original paper explicitly includes an internal quantization bottleneck. The narrower claim concerns the absence of an external discrete audio tokenizer.[5] For readers following China’s speech-model releases, that distinction helps separate a modeling choice from a slogan.

VoxCPM2 also changes the audio codec: it encodes reference audio at 16 kHz and decodes output at 48 kHz. Its language-model side operates at 6.25 steps per second of audio, retaining a rate introduced in VoxCPM1.5.[4] Those rates describe different parts of the system. A denser output waveform does not require the language model to make a separate prediction for every waveform sample.

The practical attraction, in my reading, is that richer sound and more flexible conditioning can share a relatively compact sequence. That is an architectural explanation, not a guarantee of fast response on a particular laptop.

What the release still asks of its user

The model card makes weights available under Apache 2.0 and documents streaming generation.[3] Those are useful parts of a release: developers can inspect and adapt the model, and an application can receive audio in chunks. They do not by themselves establish that an entire conversational application will feel responsive.

The usage guide draws a concrete boundary. It supports streaming audio, but says simultaneous token-by-token text input and audio generation is not supported. It recommends sentence-level handling for interactive applications. It also documents long-text failures, including speeding up, buzzing, and generation that fails to stop.[6]

For the audio-guide producer, the implication is straightforward: audition complete passages as well as attractive opening sentences. A useful trial would keep the reference recording and script fixed, change only the style instruction, and ask listeners whether the speaker still sounds like the same person. A second trial would stitch several passages together and listen for changes in pace or identity at the joins. These are suggested checks, not experiments reported here.

The important addition is a choice

VoxCPM2 broadens the family from its earlier Chinese-and-English releases to a multilingual model with explicit voice-design and controllable-cloning support.[2] Its value will depend on how reliably those controls survive ordinary scripts, accents, and editing demands.

The release nevertheless makes a useful production distinction visible. A reference recording can identify a voice; a written instruction can direct its delivery; a paired recording and transcript can anchor continuation. Returning to the imagined museum narrator, the goal is recognizable character across changing situations. VoxCPM2 gives developers clearer ways to ask for that—and clearer differences to listen for when judging the result.

Sources

1. Tsinghua University, “VoxCPM speech-generation model open-sourced,” September 25, 2025, reproducing Science and Technology Daily reporting — ModelBest and Tsinghua SIGS laboratory collaboration.
2. OpenBMB, VoxCPM Chinese-language README, commit f772e49 — April 2026 release record, supported languages and dialects, and comparison with earlier versions.
3. OpenBMB, VoxCPM2 official model card — generation modes, model access, streaming API, and license; accessed September 22, 2026.
4. Yixuan Zhou et al., “VoxCPM2 Technical Report,” June 5, 2026, version 1 — reference pathway, sequence organization, codec, and model configuration, especially sections 3.1–3.4.
5. VoxCPM Team, “VoxCPM: Tokenizer-Free TTS for Context-Aware Speech Generation and True-to-Life Voice Cloning,” September 29, 2025, version 1 — original hierarchy, internal quantization, and continuous speech generation.
6. OpenBMB, VoxCPM 2.0 Usage Guide — reference and continuation inputs, style instructions, long-text limitations, and streaming boundary; accessed September 22, 2026.
7. Karenlee9612, “Information Building at Tsinghua SIGS,” June 20, 2019 — original campus photograph and CC BY-SA 4.0 provenance, Wikimedia Commons.