c86e323d7a
Quality-first multi-pass audiobook generation from English manuscripts. Includes: - umbraaudio/ core package (pipeline, config, voices, web review UI) - scripts/ CLI utilities + model download script - tts_server/ remote TTS server - specs/ design docs, tests/, pyproject.toml + uv.lock - Qwen3-TTS git submodule (pinned 022e286) with local demo.py UI patch saved at scripts/qwen3_tts_demo_umbra_patch.patch - .env with runtime config/secrets Excluded (see .gitignore): models/ (downloaded via scripts/download_models.sh), output/, logs/, .venv, .umbra runtime data, and book projects (DealHunterShip_*), which are committed separately.
32 lines
2.2 KiB
Plaintext
32 lines
2.2 KiB
Plaintext
{\rtf1\ansi\ansicpg1252\cocoartf2870
|
|
\cocoatextscaling0\cocoaplatform0{\fonttbl\f0\fswiss\fcharset0 Helvetica;}
|
|
{\colortbl;\red255\green255\blue255;\red249\green249\blue249;\red17\green17\blue17;}
|
|
{\*\expandedcolortbl;;\cssrgb\c98039\c98039\c98039;\cssrgb\c8627\c8627\c8627;}
|
|
\margl1440\margr1440\vieww11520\viewh8400\viewkind0
|
|
\pard\tx720\tx1440\tx2160\tx2880\tx3600\tx4320\tx5040\tx5760\tx6480\tx7200\tx7920\tx8640\pardirnatural\partightenfactor0
|
|
|
|
\f0\fs24 \cf0 In the umbra audio system we are processing to validate utterances are done correctly. Chapter 6 Salvage\
|
|
Was generated and listened to. Issues became clear that need to be fixed.\
|
|
\
|
|
The system has a proper names section where the correct way to pronounce names is verified. Is that verified pronunciation used with every utterance?\
|
|
\
|
|
\pard\pardeftab720\partightenfactor0
|
|
|
|
\fs26 \cf2 \cb3 \expnd0\expndtw0\kerning0
|
|
Checking the chapter Salvage, utterance 56 isn't correct. It is actually three sentences and should be three utterances with different speakers. \
|
|
\
|
|
Based on context the utterance 145 should be modified to use a pleading tone. \
|
|
\
|
|
Utterances starting at 186 uses the Deal Hunter voice. When the narative clearly talks about a robot. The speaker should be the robot that we know from later in this section (from the context) is Nanny. \
|
|
\
|
|
Utterance 186 follow "She asks", clearly the rescued girl. From the context - Kainda has red hair, This person rescued from the yaht debre has red hair - it is Kainda. The speaker assigned is again deal hunter which is wrong. \
|
|
\
|
|
At utterance 211 suddently shifts to nanny as the speaker - this is correct, but it should have been nanny before. \
|
|
\
|
|
The utterance 218 is "<<<>>>" a section break and should be removed is is unspoken. \
|
|
\
|
|
utterance 221 speaker is nanny and correct. utterance 222 is a different person (Kainda) but the system assigns the speaker as nanny. \
|
|
\
|
|
Utterance 353 clearly has the captain yells, so utterance 354 should be yelling. Utterance 254 is correctly identified as captain willamina wake, but not yelling. \
|
|
\
|
|
These are all examples of the analysis section not working properly. The system needs to be improved. To be clear it is acceptable to have multiple passes to determine different parts of the product. This project is about quality not speed.} |