# Local validation — 2 November 2026

This record covers the supplied Maya Transcribe checkpoint. It does not reproduce the authors’ eight-condition recognition benchmark or A100/vLLM serving measurements.

## Verified

- ZIP extraction completed with CRC checks and per-file SHA256 records.
- All eight original non-card model files remain byte-identical to the archive: weights, config, generation config, processor config, tokenizer config, tokenizer, chat template and Git attributes.
- Safetensors header: 611 tensors, 782,426,112 stored elements, all BF16; 1,564,928,088 weight-file bytes.
- Native AutoProcessor and AutoModelForMultimodalLM load locally with `trust_remote_code=False`; no model/processor code was fetched.
- Real model inference passed in native and mixed styles, plus a two-clip batch with distinct prompts.
- Seven CLI checks passed: dependency-free help; stereo downmix/resampling; empty, overlong and non-finite audio rejection; existing-output preservation; invalid-style rejection.

## Environment and audio

macOS Apple Silicon, Python 3.12.14, PyTorch 2.15.0, Transformers 5.20.2, Apple MPS with float32 inference. Original BF16 weight files are unchanged. Dependencies are pinned in requirements.txt.

The smoke input is a 7.76-second, 24 kHz mono **synthetic** Hindi clip generated by Maya 2 Native, Ritvik voice. It was resampled to 16 kHz. Source: https://huggingface.co/maya-research/maya2-native/resolve/main/samples/hi_ritvik.wav . This is a functional example, not a natural-speech evaluation dataset.

## Actual outputs

**style: native**

क्या आपका वर्तमान पता पैंतालीस एमजी रोड अंधेरी ईस्ट मुंबई चार शून्य शून्य शून्य छः नौ है यदि हाँ तो एक दबाएँ

**style: mixed**

क्या आपका वर्तमान पता पैंतालीस mg road अंधेरी east मुंबई चार शून्य शून्य शून्य छः नौ है यदि हाँ तो एक दबाएँ

Both outputs terminated below the 512-token cap. Reported CLI elapsed times include loading, preprocessing and generation, and were run under different cache/warm-up conditions. They are not a latency comparison or serving benchmark.

## Not verified here

- The authors’ WER values, scorer/normalizer implementation, evaluation split independence or comparison significance.
- CUDA/A100 inference, vLLM 0.29 serving, production concurrency/latency, or minimum VRAM.
- Natural microphone recordings, noisy/phone robustness, English-only quality, diarization, timestamps, real-time streaming or long-form orchestration.

## Reproduce

```bash
pip install -r requirements.txt
python transcribe.py /path/to/authorized/hindi_clip.wav --style native
python transcribe.py /path/to/authorized/hinglish_clip.wav --style mixed
```

These commands run locally. No hosted API calls or GPU instances were provisioned in this preparation.