firdavsus/Gemma-4-E4B_hidden_to_audio_tokens-2.0

Gemma-4 S2S Alignment Dataset (Multilingual) – Version 2.0 (260K)

This dataset is specifically engineered to train a lightweight, low-latency Hidden-to-Speech (H2S) alignment model.

By capturing the raw, abstract semantic representations from the 33rd hidden states of a Text LLM (Gemma-4 8B) and mapping them directly onto quantized discrete audio streams, this dataset bypasses traditional text generation bottlenecks to establish native Speech-to-Speech (S2S) processing pipelines.

Note on Layer Selection: The 33rd layer was chosen based on empirical evidence suggesting that the 75%–80% depth layers capture a deeper semantic understanding, whereas the final layer tends to be over-specialized for next-token prediction.

Key Conceptual Blueprint

• The Paradigm: Gemma-4 (8B) natively processes multi-modal audio streaming inputs using its built-in Audio Encoder block.
• The Goal: Training a compact decoder network (e.g., an Autoregressive Language Model paired with a Non-Autoregressive acoustic upsampler) to translate Gemma's output hidden layers directly back into synchronized audio tokens. This establishes an end-to-end, low-latency voice assistant framework.

---

Dataset Structure & Specifications

The dataset is fully self-contained, serialized, and delivered in a high-throughput Apache Parquet format, optimized for memory-efficient streaming routines (IterableDataset).

• Total Scale: ~260,000 unique token sequences.

• Language Distribution: Curated in equally balanced proportions (~33% per language) across three target systems:

• 🇺🇿 Uzbek (uz)

• 🇷🇺 Russian (ru)

• 🇺🇸 English (en)

• Acoustic Profile: Rendered across a single speaker identity profile. The voice corresponds to a middle-aged female speaking fluent, standard Uzbek.

Sub-Dataset Components & Sources

The collection blends foundational conversational text prompts, translation sets, and instruction-following corpora to ensure deep linguistic variance:

Dataset Path	Focus Language	Quantity / Slices
Open-Orca/OpenOrca	English (en)	100,000 sequences
d0rj/OpenOrca-ru	Russian (ru)	100,000 sequences
sukhrobnurali/uzbek-islamic-qa-v1	Uzbek (uz)	All sequences
MLDataScientist/fleurs_En_Uz	English ⇄ Uzbek	All sequences
nickoo004/uzbekdata	Uzbek (uz)	All sequences
med-alex/qa_mt_tr_to_uzn	Turkish ⇄ Uzbek	All sequences
med-alex/qa_mt_en_to_uzn	English ⇄ Uzbek	All sequences

---

Schema & Column Breakdown

Every row vector in the Parquet file contains the following feature schemas:

Column Name	Data Type	Description
generated_text	string	The clean raw text response generated during the pipeline synthesis pass.
hidden_states_bytes	binary	Extracted float32 raw bytes representing the last hidden layer activations emitted from Gemma-4 8B.
hidden_states_shape	list(int64)	The shape matrix dimensions of the hidden states tensor (e.g., [Sequence_Length, 2560]).
codebook_bytes	binary	Quantized int32 neural audio tokens generated via the synthesis engine.
codebook_shape	list(int64)	Matrix dimension array for the discrete multi-codebook acoustic layers (e.g., [8, Audio_Frames]).

---

Data Generation Methodology

1. Semantic Extraction: Raw textual instruction configurations were fed into Gemma-4-E4B-it to generate natural, conversational text responses. A native forward hook was set on the final attention block layer to pipe out the un-padded, context-rich hidden states directly to disk.
2. Audio Synthesis & Tokenization: The text outputs were processed through the OmniVoice synthesis engine to produce high-fidelity speech.
3. Quantization Specifications: Waveform sequences were quantized using the HiggsAudioV2 neural audio tokenizer at a 24 kHz sampling rate, outputting a compact discrete matrix consisting of 8 hierarchical codebooks with a vocabulary ceiling of 1026 tokens (including boundary EOS tokens).

---

Intended Use Cases

• Stage 1 (Autoregressive Decoder): Slicing out the first structural row (codebook_bytes[0, :]) to train semantic Text-to-Acoustic alignment layers.
• Stage 2 (Non-Autoregressive Projector): Utilizing the full multi-codebook matrix to guide parallel, bidirectional acoustic upsampling models (like VALL-E or MusicGen style decoders) to synthesize high-fidelity structural vocal features.

---

Associated Code Repository

The implementation code for the pipeline, training, and models can be found in the associated GitHub repository:

https://github.com/firdavsus/Hidden-To-Speech

File Name	Description
final_dataset_old.py	The pipeline used to synthesize and generate the dataset.
read.py	Utility script to test and inspect the dataset (reads the first several rows).
model.py & NAR.py	Model architectures designed in a VALL-E style. Hyperparameters are tuned to sensible defaults, but feel free to experiment.
train.py & train_nar.py	The main training scripts for the models.