Upload folder using huggingface_hub

.gitattributes

@@ -33,3 +33,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+model.qora-stt filter=lfs diff=lfs merge=lfs -text
+qora-stt.exe filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,3 +1,257 @@
----
-license: apache-2.0
----
+---
+language:
+  - en
+  - zh
+  - de
+  - es
+  - ru
+  - ko
+  - fr
+  - ja
+  - pt
+  - tr
+  - pl
+  - nl
+  - ar
+  - sv
+  - it
+  - id
+  - hi
+  - fi
+  - vi
+  - he
+  - uk
+  - el
+  - cs
+  - ro
+  - da
+  - hu
+  - ta
+  - "no"
+  - th
+  - ur
+  - hr
+  - bg
+  - lt
+  - la
+  - ml
+  - cy
+  - sk
+  - te
+  - fa
+  - lv
+  - bn
+  - sr
+  - az
+  - sl
+  - kn
+  - et
+  - mk
+  - br
+  - eu
+  - is
+  - hy
+  - ne
+  - mn
+  - bs
+  - kk
+  - sq
+  - sw
+  - gl
+  - mr
+  - pa
+  - si
+  - km
+  - sn
+  - yo
+  - so
+  - af
+  - oc
+  - ka
+  - be
+  - tg
+  - sd
+  - gu
+  - am
+  - yi
+  - lo
+  - uz
+  - fo
+  - ht
+  - ps
+  - tk
+  - nn
+  - mt
+  - sa
+  - lb
+  - my
+  - bo
+  - tl
+  - mg
+  - as
+  - tt
+  - haw
+  - ln
+  - ha
+  - ba
+  - jw
+  - su
+license: mit
+tags:
+  - speech-to-text
+  - stt
+  - whisper
+  - rust
+  - cpu-inference
+  - pure-rust
+  - no-python
+  - no-cuda
+  - automatic-speech-recognition
+base_model: openai/whisper-tiny
+pipeline_tag: automatic-speech-recognition
+library_name: qora
+---
+
+# QORA-STT - Pure Rust Speech-to-Text
+
+Pure Rust inference engine for OpenAI's Whisper Tiny. No Python, no CUDA, no external dependencies. Single executable + binary weights = portable speech-to-text on any machine.
+
+Based on **openai/whisper-tiny** (MIT License).
+
+## Quick Start
+
+```bash
+# Transcribe an audio file (English)
+qora-stt.exe --model-path . --load model.qora-stt --audio recording.wav
+
+# Specify language
+qora-stt.exe --model-path . --load model.qora-stt --audio recording.wav --language french
+
+# Save transcription to file
+qora-stt.exe --model-path . --load model.qora-stt --audio recording.wav --output transcript.txt
+```
+
+## Files
+
+```
+model/
+  qora-stt.exe       2.5 MB    Inference engine (single binary)
+  model.qora-stt     144 MB    F32 weights (encoder + decoder)
+  config.json         2.0 KB   Model configuration
+  tokenizer.json      2.4 MB   Tokenizer (51,865 vocab)
+  README.md                    This file
+```
+
+**No safetensors needed.** Everything loads from `model.qora-stt`.
+
+## Model Info
+
+| Property | Value |
+|----------|-------|
+| **Base Model** | openai/whisper-tiny |
+| **Parameters** | 39 Million |
+| **Type** | Encoder-decoder transformer |
+| **Weights** | F32 (no quantization needed at 39M params) |
+| **Binary Size** | 144 MB |
+| **Input** | WAV audio (any sample rate, auto-resampled to 16kHz) |
+| **Output** | Transcribed text |
+| **Max Duration** | 30 seconds per chunk |
+| **Languages** | 99 languages supported |
+
+## Architecture
+
+| Component | Details |
+|-----------|---------|
+| **Encoder** | Conv1D stem (80->384, stride 2) + 4 transformer layers |
+| **Decoder** | 4 transformer layers with cross-attention to encoder |
+| **Hidden Size** | 384 |
+| **Attention Heads** | 6 (head_dim=64) |
+| **FFN Dimension** | 1,536 |
+| **Vocabulary** | 51,865 tokens (BPE) |
+| **Activation** | GELU |
+| **Normalization** | LayerNorm with bias |
+| **Mel Spectrogram** | 80 bins, n_fft=400, hop=160, 16kHz |
+| **Position Encoding** | Encoder: sinusoidal (stored), Decoder: learned |
+
+### Encoder
+1. **Conv1D stem**: Conv1(80->384, k=3, s=1) -> GELU -> Conv2(384->384, k=3, s=2) -> GELU
+2. Input: mel spectrogram `[80, 3000]` -> output `[1500, 384]`
+3. 4 transformer layers: LayerNorm -> self-attention (6 heads, full) -> residual -> LayerNorm -> FFN -> residual
+4. Final LayerNorm
+
+### Decoder (Autoregressive)
+1. Token + positional embedding
+2. 4 transformer layers, each with:
+   - Causal self-attention (with KV cache)
+   - Cross-attention to encoder output (cached once)
+   - FFN (384 -> 1536 -> 384)
+3. Output projection (tied with token embeddings)
+
+## CLI Arguments
+
+| Flag | Default | Description |
+|------|---------|-------------|
+| `--model-path <dir>` | `.` | Directory with config.json + tokenizer.json |
+| `--load <path>` | -- | Load binary model (.qora-stt) |
+| `--audio <wav>` | -- | Input WAV file to transcribe |
+| `--language <name>` | english | Language name or code (e.g., "french", "fr") |
+| `--output <path>` | -- | Write transcription to text file |
+| `--save <path>` | -- | Save binary model (for converting from safetensors) |
+| `--help` | -- | Show help |
+
+## Supported Languages
+
+99 languages including: English, Chinese, German, Spanish, Russian, Korean, French, Japanese, Portuguese, Turkish, Polish, Dutch, Arabic, Swedish, Italian, Indonesian, Hindi, Finnish, Vietnamese, Hebrew, Ukrainian, Greek, Czech, Romanian, Danish, Hungarian, Tamil, Norwegian, Thai, Urdu, Croatian, Bulgarian, Lithuanian, Latin, Malayalam, Welsh, Slovak, Telugu, Persian, Latvian, Bengali, Serbian, Azerbaijani, Slovenian, Kannada, Estonian, Macedonian, Breton, Basque, Icelandic, Armenian, Nepali, Mongolian, Bosnian, Kazakh, Albanian, Swahili, Galician, Marathi, Punjabi, Sinhala, Khmer, Shona, Yoruba, Somali, Afrikaans, Occitan, Georgian, Belarusian, Tajik, Sindhi, Gujarati, Amharic, Yiddish, Lao, Uzbek, Faroese, Haitian, Pashto, Turkmen, Nynorsk, Maltese, Sanskrit, Luxembourgish, Myanmar, Tibetan, Tagalog, Malagasy, Assamese, Tatar, Hawaiian, Lingala, Hausa, Bashkir, Javanese, Sundanese.
+
+## Performance (i5-11500, 16GB RAM, CPU-only)
+
+| Phase | Time |
+|-------|------|
+| Model Load (binary) | ~92ms |
+| Mel Extraction | ~108ms |
+| Encoder (4 layers) | ~2.6s |
+| Cross-attention Cache | ~32ms |
+| Decoding | ~26ms/token |
+| **Total (6s audio, 21 tokens)** | **~3.5s** |
+| Memory | ~144 MB |
+
+### Optimizations
+
+- **Rayon parallelism**: GEMM rows parallelized across all CPU cores
+- **Cache-friendly GEMM**: i-p-j loop order for sequential memory access
+- **Parallel attention heads**: 6 heads computed concurrently
+- **KV caching**: Cross-attention K/V computed once, reused every decoder step
+- **Self-attention cache**: Grows incrementally, no recomputation
+
+## Converting from Safetensors
+
+If you have the original `openai/whisper-tiny` safetensors:
+
+```bash
+# Download model
+huggingface-cli download openai/whisper-tiny --local-dir whisper-tiny
+
+# Convert to binary (runs one dummy transcription to trigger save)
+qora-stt.exe --model-path whisper-tiny --save model.qora-stt --audio some.wav
+```
+
+After conversion, safetensors files are no longer needed.
+
+## QORA Model Family
+
+| Engine | Model | Params | Size | Purpose |
+|--------|-------|--------|------|---------|
+| **QORA** | SmolLM3-3B | 3.07B | 1.68 GB (Q4) | Text generation, reasoning, chat |
+| **QORA-TTS** | Qwen3-TTS-12Hz | 0.6B/1.7B | 971 MB (Q4) | Text-to-speech synthesis |
+| **QORA-STT** | Whisper Tiny | 39M | 144 MB (F32) | Speech-to-text transcription |
+| **QORA-Image** | SDXS-512 | 350M | 350 MB | Text-to-image generation |
+
+All engines are pure Rust, CPU-only, single-binary executables with no Python dependencies.
+
+## License
+
+The QORA-STT inference engine is custom-built. The Whisper Tiny model weights are released under the [MIT License](https://github.com/openai/whisper/blob/main/LICENSE) by OpenAI.
+
+---
+
+*Built with QORA - Pure Rust AI Inference*

config.json

@@ -0,0 +1,144 @@
+{
+  "_name_or_path": "openai/whisper-tiny",
+  "activation_dropout": 0.0,
+  "activation_function": "gelu",
+  "architectures": [
+    "WhisperForConditionalGeneration"
+  ],
+  "attention_dropout": 0.0,
+  "begin_suppress_tokens": [
+    220,
+    50257
+  ],
+  "bos_token_id": 50257,
+  "d_model": 384,
+  "decoder_attention_heads": 6,
+  "decoder_ffn_dim": 1536,
+  "decoder_layerdrop": 0.0,
+  "decoder_layers": 4,
+  "decoder_start_token_id": 50258,
+  "dropout": 0.0,
+  "encoder_attention_heads": 6,
+  "encoder_ffn_dim": 1536,
+  "encoder_layerdrop": 0.0,
+  "encoder_layers": 4,
+  "eos_token_id": 50257,
+  "forced_decoder_ids": [
+    [
+      1,
+      50259
+    ],
+    [
+      2,
+      50359
+    ],
+    [
+      3,
+      50363
+    ]
+  ],
+  "init_std": 0.02,
+  "is_encoder_decoder": true,
+  "max_length": 448,
+  "max_source_positions": 1500,
+  "max_target_positions": 448,
+  "model_type": "whisper",
+  "num_hidden_layers": 4,
+  "num_mel_bins": 80,
+  "pad_token_id": 50257,
+  "scale_embedding": false,
+  "suppress_tokens": [
+    1,
+    2,
+    7,
+    8,
+    9,
+    10,
+    14,
+    25,
+    26,
+    27,
+    28,
+    29,
+    31,
+    58,
+    59,
+    60,
+    61,
+    62,
+    63,
+    90,
+    91,
+    92,
+    93,
+    359,
+    503,
+    522,
+    542,
+    873,
+    893,
+    902,
+    918,
+    922,
+    931,
+    1350,
+    1853,
+    1982,
+    2460,
+    2627,
+    3246,
+    3253,
+    3268,
+    3536,
+    3846,
+    3961,
+    4183,
+    4667,
+    6585,
+    6647,
+    7273,
+    9061,
+    9383,
+    10428,
+    10929,
+    11938,
+    12033,
+    12331,
+    12562,
+    13793,
+    14157,
+    14635,
+    15265,
+    15618,
+    16553,
+    16604,
+    18362,
+    18956,
+    20075,
+    21675,
+    22520,
+    26130,
+    26161,
+    26435,
+    28279,
+    29464,
+    31650,
+    32302,
+    32470,
+    36865,
+    42863,
+    47425,
+    49870,
+    50254,
+    50258,
+    50358,
+    50359,
+    50360,
+    50361,
+    50362
+  ],
+  "torch_dtype": "float32",
+  "transformers_version": "4.27.0.dev0",
+  "use_cache": true,
+  "vocab_size": 51865
+}

model.qora-stt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0cd03ac530d679e40f92141756ddb10a8ba744676405979c289c225b908af253
+size 151043268

qora-stt.exe

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:95097c10edccf113defd53bee20fb8be3b04836186fbc5ac97ad8c319001ce7b
+size 2615808