Phoneme-based language identification for Nahuatl-first evaluation in film audio.
Tenepal identifies what language is being spoken in audio, with the current public-facing evaluation focused on Nahuatl vs. Spanish. It works by analyzing the raw phoneme stream via universal phoneme recognition and matching against phonotactic profiles, with prosodic fusion as a second evidence channel. Maya support remains exploratory and is not yet presented here as a release-ready benchmark.
Tenepal — associated with Nahuatl senses around "the tongue," eloquence, or facility with words. This project name follows the
Malintzin Tenepaldiscussion summarized on the Wikipedia page for La Malinche and the scholarship cited there, especially Frances Karttunen and James Lockhart. The historical tenepal mediated between languages and cultures in conquest-era Mesoamerica.
Segment-level review of Nahuatl/Spanish predictions on La Otra Conquista. The annotator tool is in tools/annotator/.
| Metric | Value | Description |
|---|---|---|
| Hernán subset accuracy | 85.7% | Best configuration on 551 annotated NAH+SPA segments from Hernán-1-3 |
| Phoneme-only baseline | 65.7% | Same 551-segment subset, without Whisper or speaker priors |
| Cross-film LOC accuracy | 84.4% raw / 81.7% balanced | 244 annotated NAH+SPA segments from La Otra Conquista (minutes 14-44) |
| Nahuatl ASR CER | 108% → 70% | Whisper-large-v3 baseline vs. LoRA finetune on OpenSLR-92 test sample |
See PAPER.md for the full technical write-up, docs/AMITH_CORPORA.md for corpus access instructions, and EVOLUTION.md for the research journal.
Metric provenance:
85.7%,65.7%, and the551-segment Hernán subset are the current audited benchmark numbers documented in EVOLUTION.md and exported viabenchmarks/annotations/plusbenchmarks/reports/eq_comparison_gt.json.84.4% raw / 81.7% balancedcomes from the annotatedLa Otra Conquistasubset exported underbenchmarks/annotations/and discussed in PAPER.md.108% -> 70%CER is the current public draft finetuning result from PAPER.md, based on OpenSLR-92 test sampling.
pip install -e .Optional dependencies for the full pipeline:
pip install -e ".[diarization]" # pyannote.audio speaker diarization
pip install -e ".[g2p]" # epitran grapheme-to-phoneme
pip install -e ".[transcription]" # faster-whisper ASR
pip install praat-parselmouth # prosody analysis
pip install demucs # vocal isolation# Process audio file → SRT with language tags
tenepal batch input.wav --output output.srt
# Process with speaker diarization
tenepal batch input.wav --output output.srt --diarize
# Full pipeline with Whisper transcription
tenepal process input.wav --output output.srt --whisper-model small
# Live system audio capture
tenepal live
# Diagnostic: analyze phoneme distribution
tenepal analyze input.wav
# Docker GPU setup (for pyannote diarization)
tenepal setup-docker
tenepal doctorTenepal supports pluggable cloud compute for GPU-intensive pipeline stages:
# Modal (default)
modal run tenepal_modal.py::main --input audio.wav
# RunPod
TENEPAL_RUNTIME=runpod tenepal process audio.wav
# Docker (local GPU)
tenepal batch audio.wav --dockerThe runtime provider abstraction (src/tenepal/runtime/) auto-detects available backends or can be configured via TENEPAL_RUNTIME environment variable.
src/tenepal/
├── audio/ # Audio loading, format conversion, preprocessing
├── phoneme/ # Universal phoneme recognition (Allosaurus + text-to-IPA)
├── language/ # Language identification (profiles, scoring, smoothing)
├── transcription/ # Whisper integration and transcription routing
├── speaker/ # Speaker diarization (host + Docker backends)
├── prosody/ # Prosodic feature extraction (F0, rhythm)
├── fusion/ # Multi-evidence score fusion
├── preprocessing/ # Demucs vocal isolation, VAD segmentation
├── pronunciation/ # IPA → familiar-spelling rendering
├── morphology/ # Morphological analysis
├── orchestration/ # End-to-end pipeline orchestration
├── runtime/ # Cloud provider abstraction (Modal, RunPod)
├── validation/ # Scoring and evaluation framework
├── subtitle/ # SRT generation with language tags
├── docker/ # Docker GPU container utilities
├── data/ # Language profiles, epitran maps, lexicons
├── pipeline.py # Core pipeline orchestration
└── cli.py # Command-line interface
Audio → ffmpeg → Demucs (vocal isolation) → Silero-VAD (segmentation)
→ per segment:
Allosaurus → IPA → Phonotactic scoring ──┐
Parselmouth → Prosody features ──────────┤→ Score fusion → Language ID
Whisper → text (if language known) ──────┘
→ Speaker diarization (pyannote) → Speaker-level smoothing
→ SRT with [NAH|85%], [SPA|92%], [MAY|78%] language tags
The pipeline is tunable via JSON EQ config files that control scoring thresholds, fusion weights, and detection gates:
tenepal process audio.wav --eq eq_custom.jsonSee eq_default.json for all available parameters.
| Tool | Path | Description |
|---|---|---|
| Annotator | tools/annotator/ |
Web-based annotation tool for ground-truth labeling |
| Evaluation | evaluate.py + tools/corpus/ |
Scoring scripts, corpus evaluation, hallucination statistics |
| Regression | tools/regression/ |
Clip-based regression test harness |
| Corpus | tools/corpus/ |
AMITH Zacatlan-Nahuatl corpus integration |
Public-facing support docs:
- docs/AMITH_CORPORA.md — where to get the Amith corpora and how local download works
- docs/ANNOTATOR_SCREENSHOTS.md — guidance for safe public annotator screenshots
- docs/OPEN_PROBLEMS.md — current highest-priority unresolved technical problems
- PAPER.md — Technical paper: methodology, experiments, results
- EVOLUTION.md — Research journal: experiments, failures, and lessons learned chronologically
Current release emphasis:
- Hernán (2019) — the original project trigger and the main ablation benchmark; access is region-dependent and clips are not redistributed
- La Otra Conquista (1999) — the primary public-facing Nahuatl/Spanish reference, told from an indigenous perspective
- OpenSLR 92 Puebla-Nahuatl — training/evaluation corpus for Whisper finetuning
- OpenSLR 147/148 + related lexical sources — auxiliary Nahuatl corpus and lexicon sources
- Apocalypto / Maya materials — exploratory only; not yet a release-ready benchmark
Important release note: this repository does not grant rights to redistribute film audio, video, or subtitle files. The exact scenes used in experiments are documented in PAPER.md for reproduction using lawfully obtained source media.
Project framing note: Marina/Malinche remains the symbolic avatar of the project because of her role as a linguistic mediator in the conquest era, even though she does not directly appear in La Otra Conquista.
Source notes:
- Malintzin / Tenepal naming discussion: https://en.wikipedia.org/wiki/La_Malinche
- OpenSLR 92: https://openslr.org/92/
- OpenSLR 147: https://openslr.org/147/
- OpenSLR 148: https://openslr.org/148/
The Nahuatl lexicon and Whisper finetuning in this project build on speech corpora recorded and published by Jonathan D. Amith and collaborators (OpenSLR 92, 147, 148; Mozilla Common Voice Nahuatl). Without that fieldwork, none of this would work. See docs/AMITH_CORPORA.md for corpus details.
MIT License. See LICENSE.
@software{tenepal2026,
title={Tenepal: Phoneme-Based Language Identification for Endangered Languages},
author={Dresch, Markus},
year={2026},
url={https://github.com/markusd/tenepal}
}Per ardua ad astra. Built for languages the world forgot to support.