MixlyGames/PhiliaAuris-189M-Base

PhiliaAuris Model Card

PhiliaAuris is a from-scratch speech-to-text audio encoder, developed as the STT branch of the Philia project by MixlyGames (OwlNestTeam).

This model is untrained. The repository contains a complete, runnable architecture — encoder, decoder, and cross-attention are all implemented and verified to run forward/backward without errors — but the weights are randomly initialized and no training run to convergence has taken place yet. There are no WER/CER numbers because there is nothing to measure yet. This card exists to document the architecture as it stands and to track progress honestly as training begins, rather than being written retroactively once results exist.

Model Specifications

• Developer: MixlyGames (OwlNestTeam)
• Model Type: Audio Encoder + Autoregressive Text Decoder (encoder-decoder, cross-attention), for speech-to-text
• Architecture: Custom convolutional-stem audio encoder with a Transformer stack, feeding a Transformer decoder via cross-attention
• Languages (target): Russian (primary), English
• License: Apache-2.0
• Status: Architecture complete, pre-training not started

Architectural Features

• Audio front-end: Hand-implemented log-mel spectrogram (custom mel filterbank, no torchaudio dependency), batched torch.stft.
• Encoder stem: 2-layer 1D convolutional stem, stride 2 each (×4 total time downsampling) before the Transformer stack.
• Positional encoding: Sinusoidal, fixed (not learned).
• Encoder attention: Standard multi-head self-attention, nn.MultiheadAttention, pre-norm.
• Projector: 2-layer MLP projecting encoder hidden states into the decoder's cross-attention space.
• Cross-attention: Dedicated module (PhiliaCrossAttention) using scaled_dot_product_attention, query from decoder, key/value from encoder, independent hidden sizes on each side.
• Config: Fully HuggingFace-compatible (PretrainedConfig, registered with AutoConfig, supports save_pretrained/from_pretrained).

Parameters

Component	Parameter	Value
Audio front-end	sample_rate	16,000 Hz
Audio front-end	n_mels	128
Audio front-end	n_fft / win_length	400
Audio front-end	hop_length	160
Encoder	hidden_size	768
Encoder	num_layers	12
Encoder	num_heads	12
Encoder	intermediate_size	3,072
Encoder	max_audio_length	3,000 frames
Projector	hidden_size	1,024
Decoder	hidden_size	512
Decoder	num_layers	6
Decoder	num_heads	8
Decoder	intermediate_size	2,048
Decoder	max_target_length	448 tokens
Decoder	vocab_size	32,000

Parameter Breakdown

• Audio Encoder: ~89.0M parameters
• Text Decoder: ~67.6M parameters
• Projector & Embeddings: ~32.8M parameters
• Total Trainable Parameters: ~189.4M

---

Training Details

No training has been run yet. This section will be filled in with dataset composition, hardware, hyperparameters, and loss curves once pre-training starts, following the same standard this card family uses for Orpheus-Zero.

Known Issues

• Padding-mask alignment between the raw waveform sample rate and the encoder's downsampled output rate is currently being corrected — batches with variable-length audio may see incorrect masking until this lands. Flagging it here for anyone forking the code at this stage.

Quickstart

Not applicable yet — there is no usable checkpoint. The architecture can be inspected and instantiated directly:

from philia_config import PhiliaAurisConfig
from philia_auris import PhiliaAurisEncoder

config = PhiliaAurisConfig()
model = PhiliaAurisEncoder(config)

---

Roadmap and Future Work

1. Dataset preparation: Finalize and validate the audio dataset pipeline (prepare_audio_dataset.py).
2. First training run: Single-batch sanity check, then a full pre-training pass on the encoder-decoder pair.
3. First metrics: WER/CER on a held-out set, published in this card.
4. Checkpoint release: A save_pretrained-compatible checkpoint usable for inference.
5. Integration: Connect PhiliaAuris as the audio front-end for the broader Orpheus assistant pipeline.