Rename to Mesko TTS and mark checkpoint as training in progress

README.md

@@ -6,7 +6,6 @@ library_name: pytorch
 pipeline_tag: text-to-speech
 tags:
 - text-to-speech
-- voice-cloning
 - streaming-tts
 - sparse-attention
 - low-rank
@@ -17,54 +16,57 @@ datasets:
 - keithito/lj_speech
 ---
 
-# Mesko TTS Model
+# Mesko TTS
 
-Mesko TTS Model is a research-stage text-to-speech model from MesklinTech, built for fast, controllable, streaming-friendly speech generation.
+Mesko TTS is MesklinTech's dedicated text-to-speech research project.
 
-The project explores a sparse-energy architecture for TTS: low-rank projections, sparse temporal routing, laminar refinement, explicit speaker conditioning, duration/pitch/energy controls, and a compact acoustic decoder. The goal is to move toward world-class, low-latency TTS that can run efficiently and eventually support real-time voice products.
+This repository is currently published as an architecture and training-code release. The previous small checkpoint was useful for code smoke tests, but it did not include a properly trained production vocoder and produced noisy waveform audio. We have therefore marked the project as **not production-trained yet** and are continuing training before publishing a listenable release checkpoint.
 
-## Company Story
+## Mission
 
-MesklinTech is building practical AI systems from first principles, with a focus on models that are efficient, understandable, and useful outside large-lab infrastructure. Mesko TTS is part of that mission: a speech model designed around speed, controllability, and a clear path toward streaming voice generation.
+MesklinTech is building practical AI systems from first principles: compact, efficient, understandable models that can run outside large-lab infrastructure. Mesko TTS is our speech effort: a fast, streaming-oriented TTS stack designed around sparse routing, explicit acoustic control, and low-latency inference.
 
-Our long-term vision is to create a world-class TTS stack that can support real-time assistants, education tools, accessibility products, creator workflows, and voice interfaces for businesses. We are still early, but the foundation is intentionally different: compact sparse modules, explicit acoustic controls, and a design that prioritizes fast inference rather than only scaling parameter count.
+Our goal is to build a world-class fast streaming TTS system for real-time assistants, accessibility products, education tools, creator workflows, and business voice interfaces.
 
-We are actively looking for collaborators, partners, and aligned supporters who want to help build a faster, more accessible TTS future. To learn more or support the work, visit:
+We are actively looking for collaborators, partners, and aligned supporters who want to help us move from research prototype to a polished voice system. To learn more or support the work, visit:
 
 **https://mesklintech.com**
 
-## What Is Included
+## Current Status
 
-This repository contains a compact LJSpeech-trained checkpoint and the source needed to load and inspect the model.
+Status: **research / training in progress**
 
-- `model.safetensors`: model-only weights converted from local checkpoint `step_8000.pt`
-- `config.json`: Mesko/BioVoice TTS architecture and training config
-- `tokenizer.json`: tokenizer used for the LJSpeech training run
-- `training_summary.json`: checkpoint metrics summary
-- `training_metrics_step_8000.json`: metrics saved with the selected checkpoint
-- `bio_voice_tts/`: TTS model, audio features, datasets, training, inference, streaming, and vocoder code
-- `bio_llm/`: shared sparse-energy model utilities used by the project
+What is available now:
 
-The optimizer state and intermediate checkpoints were intentionally removed to keep the repository small.
+- TTS architecture source code
+- sparse semantic encoder
+- speaker encoder
+- duration, pitch, and energy predictors
+- sparse acoustic decoder
+- sparse neural vocoder code
+- LJSpeech training scripts and config structure
 
-## Checkpoint Metrics
+What is not ready yet:
 
-Selected checkpoint: `step_8000`
+- production-quality speech checkpoint
+- trained neural vocoder release
+- standardized MOS / WER / speaker-similarity benchmark
+- long-form streaming quality validation
 
-| Metric | Value |
-|---|---:|
-| loss | 1.7827 |
-| mel_loss / mel_mae | 1.6115 |
-| duration_loss | 0.0077 |
-| pitch_loss | 0.3701 |
-| energy_loss | 1.3269 |
-| speaker_cosine proxy | 1.0000 |
+## Architecture Direction
 
-These are internal training metrics from the local run. They are not a substitute for standardized MOS, WER, speaker-verification EER, or human preference evaluation.
+Mesko TTS is built around:
 
-## Architecture
+- low-rank Q/K/V projections
+- causal sparse candidate attention
+- local, memory, landmark, and content candidate routing
+- laminar excitatory/inhibitory refinement
+- explicit speaker conditioning
+- explicit duration, pitch, and energy modeling
+- compact acoustic decoding
+- streaming-oriented state/cache structure
 
-The model path is:
+The intended model path is:
 
 1. Reference mel -> speaker encoder
 2. Text tokens -> sparse semantic encoder
@@ -73,72 +75,15 @@ The model path is:
 5. Pitch and energy predictors -> frame-level controls
 6. Frame states + speaker + pitch + energy -> sparse acoustic decoder
 7. Acoustic energy/gating head -> mel spectrogram
-8. Optional sparse neural vocoder -> waveform
-
-Core design ideas:
-
-- low-rank Q/K/V projections
-- causal sparse candidate attention
-- local, memory, landmark, and content candidate routing
-- laminar excitatory/inhibitory refinement
-- explicit duration, pitch, and energy modeling
-- compact acoustic decoding
-- streaming-oriented structure
-
-## Minimal Loading Example
-
-```python
-import json
-import torch
-from safetensors.torch import load_file
-
-from bio_voice_tts import BioVoiceConfig, BioVoiceTTS
-
-def merge_dataclass(instance, payload):
-    for key, value in payload.items():
-        current = getattr(instance, key)
-        if hasattr(current, "__dataclass_fields__") and isinstance(value, dict):
-            merge_dataclass(current, value)
-        else:
-            setattr(instance, key, value)
-    return instance
-
-config = merge_dataclass(BioVoiceConfig(), json.load(open("config.json")))
-model = BioVoiceTTS(config)
-model.load_state_dict(load_file("model.safetensors"), strict=False)
-model.eval()
-
-token_ids = torch.randint(0, config.semantic.vocab_size, (1, 16))
-reference_mel = torch.randn(1, 128, config.audio.n_mels)
-
-with torch.no_grad():
-    outputs = model(token_ids, reference_mel)
-
-print(outputs["mel"].shape)
-```
-
-For waveform synthesis, pass `outputs["mel"]` through `bio_voice_tts.vocoder.sparse_vocoder.SparseNeuralVocoder`. A separately trained vocoder checkpoint is recommended for production-quality audio.
-
-## Current Status
-
-Mesko TTS Model is an early research checkpoint, not a finished production voice product.
-
-Current strengths:
+8. Trained neural vocoder -> waveform
 
-- compact checkpoint
-- fast architecture direction
-- sparse, inspectable model internals
-- explicit acoustic controls
-- source included for research and extension
+## Why The Checkpoint Was Removed
 
-Current limitations:
+The previous uploaded checkpoint could generate mel tensors, but audio generated through a fallback Griffin-Lim renderer was noisy. That is not acceptable for a public TTS release. A real TTS release needs a trained waveform vocoder or a high-quality external vocoder path.
 
-- trained on LJSpeech-style single-speaker data
-- no standardized public MOS/WER/EER benchmark yet
-- text-to-mel checkpoint; production waveform quality requires a matching vocoder
-- some configuration fields are still scaffolded for future work
+Until that is ready, this repository should be treated as source code and architecture documentation, not as a finished voice model.
 
 ## Responsible Use
 
-Do not use this model to impersonate people, clone voices without consent, commit fraud, or create misleading audio. Voice technology should be built and used with permission, transparency, and care.
+Do not use this project to impersonate people, clone voices without consent, commit fraud, or create misleading audio. Voice technology should be built and used with permission, transparency, and care.
 

model.safetensors

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:d06b15977562a2d15dd9fef7e013cea65bb147f436750f4b34fa61b85bc48a8f
-size 5835812

training_metrics_step_8000.json

@@ -1,12 +0,0 @@
-{
-  "step": 8000,
-  "metrics": {
-    "loss": 1.7827284336090088,
-    "mel_loss": 1.611472487449646,
-    "duration_loss": 0.007744799368083477,
-    "pitch_loss": 0.37013131380081177,
-    "energy_loss": 1.3269374370574951,
-    "mel_mae": 1.611472487449646,
-    "speaker_cosine": 1.0
-  }
-}

training_summary.json

@@ -1,14 +0,0 @@
-{
-  "published_checkpoint": "step_8000",
-  "step": 8000,
-  "metrics": {
-    "loss": 1.7827284336090088,
-    "mel_loss": 1.611472487449646,
-    "duration_loss": 0.007744799368083477,
-    "pitch_loss": 0.37013131380081177,
-    "energy_loss": 1.3269374370574951,
-    "mel_mae": 1.611472487449646,
-    "speaker_cosine": 1.0
-  },
-  "note": "model.safetensors contains model weights only; optimizer state was intentionally omitted to keep the Hugging Face repo small."
-}