Multilingual speech-to-text translation: speak in any supported language or a code-switched mix, and receive text in a single, pre-selected target language.
Sawti takes spoken input (including mid-utterance code-switching across supported languages) and emits a continuous text stream in a target language chosen beforehand. The initial focus is English, Arabic, and French.
Sawti is built as a composable pipeline of existing state-of-the-art models, designed so individual components can later be fine-tuned or replaced:
audio stream
│
▼
SEGMENTATION VAD · pause detection · max-duration · overlap
│ (└─ side channel: LID / confidence / diagnostics — non-blocking)
▼
S2TT ENGINE SeamlessM4T, target_lang = pre-selected
│
▼
QUALITY / FALLBACK retry · re-chunk · ASR+MT fallback behind the gate
│
▼
POST-PROCESSING merge · dedupe · punctuate · normalize per target language
│
▼
target-language text stream
- No hard LID routing. The S2TT engine infers the source language; LID runs only as a non-blocking side channel for diagnostics.
- Chunk coherence over language purity. A multi-language phrase stays one chunk — splitting around a language switch hurts translation.
- Conservative post-processing. Merge/dedupe/punctuate/normalize only. No paraphrasing, summarizing, or inference.
- Fallback behind the quality gate. Recovery is triggered by measurable quality conditions, not ad hoc handling.
- Chunked, not token-level streaming, for MVP. True streaming S2TT is a later milestone.
🚧 Early design phase. See the design specification under
docs/superpowers/specs/ for the full architecture.
uv sync
uv run pytest # run the test suite
uv run sawti transcribe --target eng # stub pipeline demo
uv run sawti eval tests/fixtures --target eng # eval harness skeletonuv sync
uv run pytest # unit suite (fast, hermetic)
SAWTI_RUN_INTEGRATION=1 uv run pytest -m integration # real-model tests (GPU, ~2.3GB download)
uv run sawti transcribe sample.wav --target eng --engine m4t # real pipeline--engine m4t loads SeamlessM4T-v2-large (CUDA) and requires a real audio
file. Without a file, or with --engine stub, the stub pipeline runs
(hermetic, no model download).
M1 acceptance note: the unit suite (hermetic) is the CI bar. The real
SeamlessM4T integration test and the --engine m4t smoke are implemented
but deferred pending first-time model download and a real speech sample:
SAWTI_RUN_INTEGRATION=1 uv run pytest -m integration
uv run sawti transcribe path/to/speech.wav --target eng --engine m4tTo be determined.