magic-linguistic-speech
When to Use
- Ingesting field-recorded audio + linguistic annotation (ELAN, Praat, FLEx, SayMore) into ML pipelines.
- Choosing G2P (grapheme-to-phoneme) approach for the target language.
- Selecting low-resource ASR (MMS / Whisper / fine-tune).
- IPA validation in transcription pipelines.
- Building Lhotse CutSet from heterogeneous community-annotated sources.
- Bridging endangered-language oral data into TTS / ASR research.
When NOT to use: purely text data → not speech-relevant. Pure tokenizer audit → magic-linguistic-tokenize. Annotation methodology → magic-linguistic-annotate.
The Knowledge Engineers Routinely Miss
-
ELAN tier-naming conventions vary across community projects. One project uses tier "ipa" for IPA transcription; another uses "phonetic"; another "tx@en"+lang-suffix. Standardize at ingest, not at downstream consumption.
-
FLEx FieldWorks XML often uses SIL PUA (Private Use Area) characters from older fonts. Pre-Unicode legacy. Convert to Unicode at ingest or downstream tools choke silently.
-
Lhotse CutSet (a "cut" = audio + supervisions + features) is the standard 2026 representation. ESPnet, k2/icefall, SpeechBrain, NeMo all consume it. Build pipelines that produce CutSets, not bespoke formats.