Skip to content

v1.1 — Built-in TTS Engine & Improved Script Quality

Choose a tag to compare

View all tags
@Finrandojin Finrandojin released this 08 Feb 01:35
· 152 commits to main since this release
v1.1
1d51840

Alexandria v1.1 — Built-in TTS Engine & Improved Script Quality

Two major areas of improvement: a fully integrated local TTS engine that eliminates the need for an external server, and a significantly better LLM prompt that produces more faithful, expressive audiobook scripts.

Built-in Qwen3-TTS Engine

  • No external server required — TTS models load directly in-process, no separate Gradio server needed
  • Batch (Fast) rendering — New render mode sends multiple lines to TTS in a single batched call for 3-6x real-time throughput
  • Codec compilation — Optional torch.compile optimization for 3-4x faster batch decoding (toggle in Setup tab)
  • Smart sub-batching — Automatically groups similarly-sized text chunks to minimize wasted GPU compute on padding
  • Configurable from UI — Sub-batch settings (enabled, min size, length ratio) and compile toggle exposed in Setup tab
  • ROCm auto-optimizations — MIOpen fast-find mode, Triton flash attention, and triton_key compatibility shim applied transparently on AMD GPUs

Improved Script Generation Prompts

  • Hearing test rule — Narrator/character split now uses "would another character in the scene hear it?" instead of the old first-person heuristic that was converting third-person narration into character dialogue
  • Author text preservation — Explicit rules to never change person, tense, or wording. The LLM no longer rewrites "His experience was like living in the past" as a character saying "My experience is like living in the past"
  • Attribution tag handling — Dialogue tags like "said Brin" are dropped (voice assignment replaces them), but descriptive actions from attributions are preserved as NARRATOR text
  • Richer instruct directions — Changed from "3-8 words" to "2-3 sentences" for dramatically better TTS emotiveness and prosody
  • Better context continuity — Each chunk now receives a character roster and the last 3 script entries instead of the broken "main character" frequency heuristic

Character Style (replaces Default Style)

  • New character_style field — Persistent character traits (accents, speech patterns) that append to every TTS instruct for that speaker
  • Examples: "Heavy Scottish accent", "Refined aristocratic tone", "Speaks in clipped military cadence"
  • Backward compatible with existing default_style in voice configs

Other Changes

  • Cleaned up default prompts for general audience
  • Fixed triple-quote syntax error in app.py DEFAULT_SYSTEM_PROMPT
  • Updated README with new API examples, instruct style guidance, and performance documentation
  • Added audiobook sample to README

Performance

Tested on AMD RX 7900 XTX (24 GB VRAM, ROCm 6.3):

Configuration Throughput
Standard mode (sequential) ~1x real-time
Batch mode, no codec compile ~2x real-time
Batch mode + compile_codec 3-6x real-time

A 273-chunk audiobook (~54 minutes of audio) generates in approximately 16 minutes with batch mode and codec compilation enabled.

Assets 2
Loading