YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

VoxCPM2 Burmese Mobile — VXM v2

Golden Reference Build

VXM v2 is a verified, self-contained reference build for the VoxCPM2 Burmese Mobile Runtime.

Python = Truth

The build was reconstructed and verified against the canonical Python VoxCPM2 model without loading the full model.safetensors during the final clean reconstruction.

Quick Release Summary

Why is VXM v2 2.68 GiB?

VXM v1 was approximately 1.70 GB because it contained the original core runtime files.

During the canonical Python parameter-provenance audit, VXM v1 was found to be missing 198 parameters required for a complete clean reconstruction:

Component	Missing parameters
Base LM	58
Residual LM	17
`feat_decoder`	123
Total	198

These 198 parameters were confirmed in the canonical model.safetensors source and were extracted into VXM v2.

VXM v2 also contains the corrected canonical enc_to_lm_proj weights because the original VXM copy did not match the canonical Python model.

Therefore the package increased from approximately 1.70 GB to 2.68 GiB.

The additional size is intentional.

It makes VXM v2 substantially more self-contained for the verified Python reconstruction instead of depending on the original full model.safetensors file.

What Was Added to VXM v2?

VXM v2 contains the original runtime files plus additional parameter packages required for the clean reconstruction.

Original VXM extras

feat_encoder_bf16.safetensors
fsq_bf16.safetensors
fusion_bf16.safetensors
projection_bf16.safetensors
stop_head_bf16.safetensors

Corrected / newly added extras

enc_to_lm_proj_bf16.safetensors
base_missing_bf16.safetensors
residual_missing_bf16.safetensors
feat_decoder_bf16.safetensors

Total extra tensors:

325 tensors

Why Was `enc_to_lm_proj` Corrected?

During clean reconstruction, the original VXM enc_to_lm_proj_bf16.safetensors was compared against the canonical Python model.

The values did not match.

The original VXM copy produced incorrect audio during the clean reconstruction.

The canonical enc_to_lm_proj weight and bias were therefore extracted directly from the canonical model.safetensors source and used to replace the incorrect VXM copy.

After the correction:

Clean reconstruction       : PASS
Mapping failures            : 0
META parameters             : 0
enc_to_lm_proj              : PASS
Canonical audio generation  : PASS

The corrected VXM v2 reconstruction produced the expected audio.

How Was the Python Reconstruction Verified?

The final clean reconstruction did not load the full model.safetensors.

It used:

Base LM INT4
        +
Residual LM INT4
        +
VXM original extras
        +
198 missing parameters
        +
Corrected enc_to_lm_proj
        ↓
Fresh VoxCPM2Model
        ↓
Canonical Python generation

Canonical test:

အားလုံးကျန်းမာချမ်းသာကြပါစေ

Generation settings:

inference_timesteps = 5
cfg_value            = 2.5
seed                 = 1234

Result:

Sample rate : 48000 Hz
Samples     : 84480
Duration    : 1.76 seconds
dtype       : float32

The resulting audio was verified as correct.

Python reconstruction: PASS

Status

VXM v2 = GOLDEN REFERENCE BUILD

This release is the verified reference build for the VoxCPM2 Burmese Mobile Runtime.

Verification Status

Python reconstruction: PASS
Parameter provenance: PASS
Clean reconstruction without model.safetensors: PASS
enc_to_lm_proj canonical correction: PASS
Audio generation: PASS
Audio output verification: PASS
VXM binary format: PASS
Embedded entry verification: PASS
Embedded SHA256 verification: PASS

1. Project Goal

The target runtime pipeline is:

Burmese Text
    ↓
VoxCPM2 Mobile Runtime
    ↓
Full LM Generation
    ↓
DIT CFM
    ↓
AudioVAE Decoder
    ↓
48 kHz Burmese WAV

The Python VoxCPM2 implementation is the canonical source of truth.

Python = Truth

Future C++ and mobile implementations must reproduce the canonical Python behavior.

2. Why VXM v2 Was Created

The original VXM v1 contained the main runtime files.

A complete provenance audit showed that the VXM v1 package did not contain all parameters required to reconstruct the canonical Python model exactly.

The audit identified:

Base LM missing parameters: 58
Residual LM missing parameters: 17
AudioVAE feat_decoder parameters: 123

Total additional parameters:

198

All 198 parameters were confirmed to exist in the original model.safetensors.

Therefore these parameters were extracted from the canonical source and added to the VXM v2 package.

3. `enc_to_lm_proj` Correction

During clean reconstruction, the original VXM extra enc_to_lm_proj_bf16.safetensors was found to contain values that did not match the canonical Python model.

The canonical enc_to_lm_proj weight and bias were extracted directly from the original model.safetensors.

The corrected component was then used for the VXM v2 reconstruction.

Result

Before correction:

Audio output: incorrect

After canonical correction:

Audio output: correct

This correction was required for exact Python audio parity.

4. VXM v2 Extra Weights

VXM v2 contains the original required extras plus the missing parameters discovered during provenance analysis.

Existing extras

feat_encoder_bf16.safetensors
fsq_bf16.safetensors
fusion_bf16.safetensors
projection_bf16.safetensors
stop_head_bf16.safetensors

Corrected / additional extras

enc_to_lm_proj_bf16.safetensors
base_missing_bf16.safetensors
residual_missing_bf16.safetensors
feat_decoder_bf16.safetensors

Total extra tensor count:

325 tensors

5. Clean Reconstruction

A fresh VoxCPM2Model was reconstructed without loading the full model.safetensors.

The reconstruction used:

Base LM INT4 weights
Residual LM INT4 weights
VXM extra weights
198 additional missing parameters
Corrected canonical enc_to_lm_proj

Verification

Mapping failures      : 0
META parameters       : 0
Parameter structure   : PASS
AudioVAE parameters   : PASS
enc_to_lm_proj        : PASS

The clean reconstruction successfully generated the expected audio.

6. Canonical Audio Verification

Canonical test text:

အားလုံးကျန်းမာချမ်းသာကြပါစေ

Canonical generation:

text = "အားလုံးကျန်းမာချမ်းသာကြပါစေ"

with torch.inference_mode():
    wav = model.generate(
        target_text=text,
        inference_timesteps=5,
        cfg_value=2.5,
        seed=1234,
    )

Expected output:

Sample rate : 48000 Hz
Samples     : 84480
Duration    : 1.76 seconds
dtype       : float32

VXM v2 clean reconstruction produced the correct audio.

Audio verification: PASS

7. VXM Container Format

VXM v2 uses a custom binary container format.

Magic:
VOXCPM2\0

Version:
2

Entry count:
18

Each entry contains:

filename
data offset
data size
reserved space

VXM v2 is not a ZIP file.

The file begins with the VXM magic:

VOXCPM2\0

8. VXM v2 Contents

Core runtime files

manifest.json
base_lm_int4.pt
residual_lm_int4.pt
voxcpm2_dit_fp16.onnx
audiovae.pth
tokenizer.json
tokenizer_config.json
special_tokens_map.json
config.json

Extra weights

extras/base_missing_bf16.safetensors
extras/enc_to_lm_proj_bf16.safetensors
extras/feat_decoder_bf16.safetensors
extras/feat_encoder_bf16.safetensors
extras/fsq_bf16.safetensors
extras/fusion_bf16.safetensors
extras/projection_bf16.safetensors
extras/residual_missing_bf16.safetensors
extras/stop_head_bf16.safetensors

Total entries:

9. AudioVAE Encoder Investigation

Runtime tracing showed that the normal text-to-speech generation path executes the AudioVAE decoder.

The AudioVAE encoder contains 121 parameters that were not directly executed by the normal generation path.

A controlled trap test replaced the AudioVAE encoder with a zero-parameter trap that would fail if called.

Generation still succeeded.

The generated waveform remained:

Shape : (1, 84480)
dtype : float32

Exact comparison with the baseline:

torch.equal      : True
max abs diff     : 0
mean abs diff    : 0
nonzero diff     : 0

Therefore:

The AudioVAE encoder is not required for the current normal text-to-speech generation path.

The encoder weights are intentionally retained in VXM v2 because they may be required for future reference-voice / voice-cloning functionality.

10. Future Runtime Optimization

VXM v2 is intentionally preserved as the golden reference.

Future versions may investigate:

lazy-loading unused parameters
separating AudioVAE encoder and decoder weights
reducing startup RAM usage
reducing package size
predefined reference voices
voice-cloning runtime optimization

Any optimization must preserve the canonical Python behavior.

Future optimized builds should use a new VXM version rather than modifying this golden build.

11. Golden Build Information

File

voxcpm2_mobile_v2.vxm

Size

2,882,753,940 bytes

SHA256

0722548355a55cf485e44850e7a9e0b711fb3de5eeefd1a1f7d2393cdeb80740

Container

Magic       : VOXCPM2\0
Version     : 2
Entry count : 18

12. Golden Reference Rule

VXM v2 is the immutable reference build.

Do not modify or replace this file during future optimization work.

Future builds should use new versions:

VXM v3
VXM v4
...

If a future optimization fails, VXM v2 remains the known-good rollback reference.

13. Final Status

Python reconstruction       PASS
Parameter provenance        PASS
Clean reconstruction        PASS
enc_to_lm_proj correction   PASS
Audio generation            PASS
Audio output                PASS
VXM binary format           PASS
Entry verification          PASS
SHA256 verification         PASS
AudioVAE decoder            REQUIRED
AudioVAE encoder            RETAINED / LAZY-LOAD CANDIDATE

Golden Release

VXM v2 is the verified Golden Reference Build for the VoxCPM2 Burmese Mobile Runtime.

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support