ottis

By the numbers

faster than real time

~4×

4.96 s of speech in 1.19 s on a Ryzen 9 9900X3D, 4 threads

from word cue to the voice

~30 ms

median +10 ms against the onset after a pause

binary, nothing to install

678 KB

the runtime, model and aligner download on first use

of real time to align

2%

0.47 s for 21.4 s of audio

Speed from ottis say (2026-10-12, voice george, seed 12); alignment numbers from PERFORMANCE_TRADEOFFS.md, measured against pause onsets and Whisper medium.en word timestamps.

What's in it

1. 01Pocket TTS, no Python

   Kyutai's 100M-parameter Pocket TTS runs through ONNX Runtime's C API, loaded at run time, with a SentencePiece tokenizer and the chunking and generation loop written in Odin.

2. 02A time for every word

   Each chunk's text is force-aligned to its audio by a wav2vec2 CTC recognizer at 20 ms frames, so cues land where the word is actually spoken, not where the model's attention puts it.

3. 03Honest about numbers

   Words the aligner can't spell, like "$3.50" said as "three fifty", share the span between their neighbours and are marked estimated.

4. 04Same seed, same samples

   The only randomness is the flow head's noise from ottis's own seeded generator, and ONNX Runtime's thread count is fixed, so a seed reproduces its audio bit for bit; say prints the SHA-256.

5. 05Downloads, pinned

   ONNX Runtime 1.30.0, the english_2026-04 bundle, the voice and the aligner are fetched on first use and checked against SHA-256 pins, over libcurl on Linux and macOS and WinHTTP on Windows.

6. 06Whisper-style JSON

   The -words output is the shape Whisper writes, so caption tools read it as is; onimite's captions{} does.

7. 07A library too

   After engine_init the engine is read-only: synthesize runs on several threads at once and gives the same samples as alone. pocket.VERSION changes whenever the output would.

8. 08Ottis, who lip-syncs

   The mascot is an SVG drawn by tools/mascot with eyes, four mouths and sound waves as layers a script switches by id, so Ottis can mouth ottis's own word cues.

Use it

odin build . -o:speed -out:build/ottis
build/ottis say "The lamp cost $3.50." -o lamp.wav -words lamp.json
build/ottis say -f script.txt -voice marius -seed 3 -o scene.wav -words scene.json
build/ottis voices        # the 27 voices
build/ottis fetch -verify # download up front, or re-hash what is there

{"text": "The lamp cost $3.50.", "duration": 1.840, "segments": [
  {"id": 0, "start": 0.000, "end": 1.840, "text": "The lamp cost $3.50.", "words": [
    {"word": "The", "start": 0.162, "end": 0.222},
    {"word": "lamp", "start": 0.283, "end": 0.485},
    {"word": "cost", "start": 0.586, "end": 0.950},
    {"word": "3", "start": 1.132, "end": 1.220, "estimated": true},
    {"word": "50", "start": 1.220, "end": 1.395, "estimated": true}]}
]}

import "deps:ottis/assets"   // with -collection:deps=.. from a sibling project
import "deps:ottis/pocket"

root, _ := assets.data_dir()
library, err := assets.fetch(root, "alba")          // runtime, bundle, aligner, voice
engine: pocket.Engine
pocket.engine_init(&engine, root, library)           // once per process
voice, _ := pocket.voice_load(&engine, root, "alba") // once per voice
result: pocket.Result
pocket.synthesize(&engine, &voice, "Hello there.", pocket.DEFAULT_OPTIONS, &result)
// result.audio: mono f32 at result.sample_rate (24 kHz)
// result.words: { text, index, start, end, estimated } for every source word

The family

• ovg2D vector graphics: paths, SVG, Lottie, text, filters
• odecsVideo, audio and image codecs with no FFmpeg
• okslSkSL and Shadertoy shaders to SPIR-V, MSL and CPU
• ottisText to speech with a time for every word
• onimiteVideos from Lua scripts, written by you or your coding agent