Skip to content
This repository was archived by the owner on Sep 21, 2026. It is now read-only.

v1.5.8 — five months since v1.4.2

Latest
Latest

Choose a tag to compare

View all tags
@LattifaiHQ LattifaiHQ released this 27 Apr 13:58
· 40 commits to main since this release
v1.5.8
19d6dbf

Highlights since v1.4.2

This is the cumulative summary of every release shipped between v1.4.2 (2026-02-26) and v1.5.8 (2026-04-26) — five months of work covering a major config overhaul, a brand-new auth system, the full translation / diarization / summarize pipelines, vLLM and MLX inference backends, ten new CLI commands, and the YouTube + ASS karaoke rewrites.

Heads up: v1.5.0 contained breaking changes to CaptionConfig and CLI flags. If you are upgrading directly from v1.4.x, read the migration table below before bumping.

Breaking changes (v1.5.0)

CaptionConfig was flattened-into-nested. CLI dotted paths follow the same shape:

Before (v1.4.x) After (v1.5.x)
CaptionConfig(split_sentence=True) CaptionConfig(input=CaptionInputConfig(split_sentence=True))
CaptionConfig(word_level=True) CaptionConfig(render=RenderConfig(word_level=True))
CaptionConfig(karaoke=KaraokeConfig(enabled=True)) CaptionConfig(ass=ASSConfig(karaoke_effect="sweep"))
caption.write(path, word_level=…, karaoke_config=…) caption.write(path, render=…, format_config=…)
caption.word_level=true (CLI) render.word_level=true (CLI)
caption.karaoke.enabled=true (CLI) ass.karaoke_effect=sweep (CLI)
AlignmentConfig(flush=N) AlignmentConfig(flush_interval=N)
lai alignment youtube … lai youtube align …
lai translate run … lai translate caption …

New CLI commands (10+)

lai auth login | logout | trial | whoami        # device-flow auth + 14-day trial
lai doctor                                       # diagnostics with bundled selftest data
lai update                                       # detects stale editable installs
lai config set KEY=VALUE                         # ~/.config/lattifai/config.toml
lai serve                                        # local 4-tab playground (align / transcribe / convert / translate)
lai summarize caption                            # chapter-based summaries with metadata, SEO, quotes
lai youtube run | align                          # one-shot YouTube → aligned captions
lai translate caption | youtube                  # full multi-provider translation pipeline
lai diarize naming                               # LLM-based speaker name inference

Top-level lai --version, --extra-index-url autoconfig via uv tool install, and a lai umbrella command replace the older lai-* scripts (the legacy entry points still work).

Authentication

  • New device-flow login (lai auth login) — opens browser, listens on a random local port, saves token to ~/.config/lattifai/config.toml
  • lai auth trial — no-signup quick start with auto-provisioned 14-day trial key
  • X-Device-Auth header auto-injection for device-bound API keys
  • lai auth whoami shows account email + plan
  • v1.5.8 adds a one-shot pre-flight warning before backend-bound commands when the trial is expired or expiring within 2 days, so you see the issue before the server returns 401

Transcription

  • vLLM / SGLang backend for Voxtral Realtime (WebSocket), Gemma-3n, Fun-ASR-Nano
  • Qwen3-ASR local model support (vendored, decoupled from transformers 5.x conflict)
  • MLX backend (Apple Silicon): MLXTranscriber with dual-backend (mlx-audio + mlx-vlm); 30-second chunk cap for backend reliability
  • Google Gemma-4 E2B / E4B support; new Gemini 3 Pro / 2.5 Flash / 2.5 Flash-Lite
  • Batch concurrent transcription for VAD chunks; dual API mode (transcriptions / chat); event detection; verbose_json
  • Auto-inject ASR system prompt and temperature for general-purpose LLMs
  • In-memory MP3 chunks, configurable max_retries (default 2), system_prompt / audio_content_type / verbose for vLLM chat mode

Translation pipeline

  • Multi-provider (Gemini / OpenAI / OpenAI-compatible) end-to-end pipeline
  • Three modes: quick (direct), normal (analyze → translate), refined (analyze → translate → review → polish)
  • Glossary management, bilingual output with translation_first ordering
  • Brand-name and no-meta-commentary rules in prompts
  • Retry with exponential backoff and checkpoint/resume for long documents

Speaker diarization

  • LLM-based speaker name inference using YouTube metadata and dialogue context
  • Candidate extraction with voting, talk-format detection, post-LLM audience correction
  • lai diarize naming — dedicated command for speaker identification

YouTube

  • 4+ transcript extraction strategies (Dwarkesh, Substack, podscripts.co, Rescript API)
  • Chrome CDP and Safari fallbacks for SPA transcript pages
  • SSL hijack detection, auto-detect caption language from video metadata
  • Parent-channel metadata resolution + reordered speaker palette

Summarize (new in v1.5.5+)

  • Chapter-based summaries with metadata, SEO fields, and notable quotes
  • honor_meta_chapters hard constraint — preserves user-supplied chapters verbatim
  • json-repair for robust LLM JSON parsing; retry on JSON failure with raw output logging
  • v1.5.8: SummarizationConfig.llm now reads [summarization] in config.toml and falls back to the SUMMARIZATION_MODEL_NAME env var — point summarization at any OpenAI-compatible endpoint (SiliconFlow, vLLM, …) without code edits. Default stays Gemini.

Alignment

  • Streaming flush-every-N-chunks for O(chunk) memory usage on long files
  • Duplicate text-block detection with diff-aware detokenization (handles repeated subtitle segments across flush boundaries)
  • Lyrics-aware: auto-suppresses duplicate output for lyrics input; G option to keep genuine duplicate blocks
  • Currency / percent / thousands aware tokenization
  • RMS volume normalization before ONNX inference
  • start_margin / end_margin unified default of 0.10s
  • AudioData.stats() + CLI for amplitude distribution
  • Preserve input Supervision.id when split_sentence=False

ASS / Karaoke

  • 12 karaoke color schemes (azure-gold, sakura-purple, mint-ocean, …)
  • 10-color speaker palette with ass.speaker_color=auto
  • ASSConfig self-contained: font, colors, outline, shadow, positioning, karaoke effect
  • kinetic_style surfaced in caption convert with auto-enable for word-level output

LLM module

  • Unified lattifai.llm: BaseLLM, GeminiLLM, OpenAICompatLLM
  • Shared LLMConfig with reasoning toggle and section-based auto-resolution
  • Gemini routed through OpenAI-compatible endpoint via shared client

Config system

  • config.toml auto-resolution: CLI defaults sourced from ~/.config/lattifai/config.toml
  • Structured CaptionConfig with sub-configs: CaptionInputConfig, CaptionOutputConfig, RenderConfig
  • Per-format configs: ASSConfig, TTMLConfig, FCPXMLConfig, PremiereConfig, LRCConfig
  • Nested TOML sections for per-module LLM configs
  • TOML round-trip preserves comments (tomlkit)

Notable fixes

  • VTT speaker prefix preserved on consecutive same-speaker cues (was deduped, now matches user expectation that explicit speakers are emitted unconditionally)
  • Auth/quota errors propagated correctly from alignment API instead of wrapping as LatticeDecodingError
  • Multilingual tokenizer fixes; label-aware boundary merge; punctuation-inflated duplicate filter
  • CJK support in align_timestamps_from_ref
  • Numpy array wrapping for Qwen3-ASR; verbose_json fallback; correct Supervision timing
  • Theme: dark orange warn for white-background terminal readability
  • 15 bugs from v1.5.0 third-party install test report (config flags, README examples, classifier mismatches, doctor false positives, lai --version)
  • 6 additional bugs from v1.5.2 install test report

Dependencies

  • lattifai-core: >=0.7.5 (was >=0.6.0 at v1.4.2)
  • lattifai-captions: >=0.4.11 (was >=0.2.2 at v1.4.2; new RenderConfig API at 0.4.0)
  • lattifai-run: >=1.0.4
  • lattifai-auth: >=0.2.1 (new)
  • k2py: 0.4.0 (upgraded from 0.2.4)
  • Python 3.10 – 3.14 supported (added 3.13 / 3.14)
  • Replaced deprecated cgi.FieldStorage with stdlib-only multipart parser (Python 3.13 ready)
  • 23k lines of dead vendored qwen_asr code removed

Install

# CLI tool (recommended)
uv tool install "lattifai[all]" --extra-index-url https://lattifai.github.io/pypi/simple/

# or pip
pip install "lattifai[all]" --extra-index-url https://lattifai.github.io/pypi/simple/

Full version history

For per-version detail see CHANGELOG.md:

  • v1.5.8 — Trial-expiry pre-flight warning · VTT speaker prefix preservation · Summarization OpenAI-compat fallback
  • v1.5.7 — Lyrics-aware dedup · currency tokenization · AudioData.stats()
  • v1.5.6 — honor_meta_chapters summarize constraint · Supervision.id preservation
  • v1.5.5 — Chapter-based summarize · brand-name translation rules · json-repair
  • v1.5.4 — MLX backend (Apple Silicon)
  • v1.5.2 — X-Device-Auth header injection · 23k-line qwen_asr cleanup
  • v1.5.1 — 15-bug fix pass from v1.5.0 install report
  • v1.5.0 — Breaking: CaptionConfig overhaul · auth · translation · diarization · YouTube · summarize · CLI rewrite

Full diff: v1.4.2...v1.5.8

Assets 2
Loading