MAJOR FIX: Switch from conversation model to dedicated STT model

BREAKTHROUGH: Fix PAD token issue with correct STT start token

Enable audio boosting for quiet speech (RMS < 0.12)

Remove cross-attention from v0_1_vision config for Q8 model

Use v0_1_vision() config for exact Q8 model dimension match

Fix Q8 model shape mismatch - let GGUF auto-configure dimensions

Fix GGUF model loading - HeaderTooLarge error resolved

Fix transformer config field names (validated) - v1.6.4

Fix LM config compilation error - v1.6.3

Fix model dimensions + pre-download moshiko-candle-bf16 models - v1.6.2

Implement exact official Moshi backend configuration - v1.6.0

Switch to kyutai/moshiko-candle-q8 quantized model - v1.5.0

Fix Rust borrowing error in bounds checking - v1.4.40

Add defensive bounds checking for Moshi lm_state.step() - v1.4.39

v1.4.38: CRITICAL FIX - Moshi dual-stream configuration for STT

DEFINITIVE FIX: Configure Mimi for exactly 8 codebooks to match STT v0_1_one_way() v1.4.18

Use proper STT-only configuration: v0_1_one_way()

Fix array bounds: limit codes to 8 elements instead of 16

Force Rust binary rebuild with v1.4.13 - includes Moshi panic fix

Fix codes mutability for truncate - make codes mutable

Fix Rust compilation warning: remove unused mut

Fix critical Moshi panic: correct audio codebook configuration

v1.4.8: Add language conditioning for multilingual STT model

v1.4.7: BREAKTHROUGH - Fix extremely quiet audio causing pad tokens

v1.4.6: SYSTEMATIC DEBUGGING - Audio analysis + step limit fix

v1.4.5: Follow moshi library patterns - Remove competing normalization

v1.4.4: CRITICAL FIX - Audio normalization destroying speech recognition

v1.4.3: Debug token filtering - Show all generated tokens

CRITICAL FIX: Resolve index out of bounds panic in audio processing

CRITICAL FIX: Correct vocab size mismatch in 1B model configuration

REVERT: Switch back to 1B multilingual model for T4 GPU compatibility

Fix vocab_size scope error by adding to MoshiAsrModel struct

Fix Rust compilation errors in model.rs

MAJOR: Switch to 2.6B English STT model to match unmute.sh

Fix lm_generate_multistream::State constructor

Fix STT implementation to use proper moshi-backend patterns

Add debugging for LM model text generation

Fix all candle_core references to use renamed candle crate

Fix candle dependency using moshi pattern

Fix candle imports to match moshi asr.rs patterns

Implement A+ plan: Real text generation from LmModel

🚀 MAJOR: Pre-load models into Docker image for instant startup

REVERT: Restore to last known working state (pre-load_streaming)

Fix compilation errors with moshi 0.6.3 API

Use official moshi-backend loading pattern: load_streaming()

Fix audio tensor rank: 2D → 3D for Mimi encoder compatibility

Fix moshi Config field names: vocab_size_in → text_in_vocab_size

Fix model vocab size mismatch: 48001 → 8001 tokens

Fix CUDA compatibility for T4 GPU by using F32 instead of BF16

Fix root cause: eliminate multiple process startup v0.8.0