Skip to content
Bodéwadmimwen ASRLive
01 — Introduction

Automatic speech recognition for the Potawatomi language

A Whisper-small model adapted with LoRA on under ten hours of community audio, built with the Citizen Potawatomi Nation Language Department.

This site is reference material: how the model works, how to use it, what it cannot do, and the terms under which the community's language data is held.

02

Overview

Bodéwadmimwen (Potawatomi) is an Algonquian language of the Great Lakes with a small and largely elder population of fluent speakers. Speech recognition matters here because transcription is the bottleneck in language work: recordings accumulate faster than they can be written down, and every untranscribed hour is documentation that cannot be searched, taught, or studied.

This is a low-resource setting in the strictest sense. Fewer than ten hours of audio are available for training, against the tens of thousands of hours behind high-resource systems. The approach therefore adapts a multilingual foundation model rather than training from scratch: Whisper-small provides the acoustic prior, and LoRA adapters specialize it on community recordings without overwriting what the base model knows.

Data care
The training corpus is held by the Citizen Potawatomi Nation. Nothing on this site publishes raw recordings, and no restricted or ceremonial material is included in released data.
03

Quick facts

0.0%
Word error rate
0.0%
Character error rate
< 0 hrs
Training audio
LoRA
Adaptation method
Whisper-small
Base model
Live
Deployment status
Transcribe demo — placeholder
Press Transcribe to see simulated output.

Illustrative only — not live model output

How to read WER and CER
Word error rate counts word-level substitutions, insertions, and deletions against a reference transcript; character error rate does the same over characters. For a polysynthetic language, where a single word carries what English spreads over a clause, the gap between 38.6% WER and 11.2% CER is expected: most errors are partial, inside long words, rather than wholly wrong words.
04

Quickstart

Send a single-channel WAV file to the hosted endpoint and read the transcription off the JSON response.

python
import requests

with open("sample.wav", "rb") as f:
    response = requests.post(
        ASR_ENDPOINT,                      # TODO: endpoint URL
        headers={"Authorization": f"Bearer {ASR_TOKEN}"},
        files={"audio": f},
        timeout=60,
    )

response.raise_for_status()
print(response.json()["text"])
Input requirementValue
FormatWAV (PCM)
ChannelsMono
Sample rateTODOconfirm
Maximum durationTODOconfirm
AuthenticationTODObearer token scheme to confirm
05

Architecture

The system is Whisper-small — an encoder-decoder transformer trained on multilingual speech — adapted with Low-Rank Adaptation (LoRA). LoRA freezes the base weights and trains small rank-decomposition matrices injected into selected projection layers, so only a fraction of a percent of parameters are updated.

Why LoRA over full fine-tuning. With under ten hours of audio, full fine-tuning overfits quickly and erodes the multilingual acoustic representation that makes the base model useful in the first place. LoRA keeps that representation intact, trains on modest hardware, and produces adapter weights small enough for the community to hold, version, and revoke independently of the base checkpoint.

LoRA configurationValue
Rank (r)TODO
AlphaTODO
DropoutTODO
Target modulesTODO
TODO
LoRA configuration must be filled in from the training run before this page is treated as authoritative. Nothing here is inferred.
06

Data & corpus

The training audio comes from the Citizen Potawatomi Nation Language Department. It is community-held data used under partnership, not a scraped or open corpus, and it is not redistributed by this project.

PropertyValue
Total audioUnder 10 hours
SourceCPN Language Department recordings
Sample rateTODO
SpeakersTODO
Train / validation / test splitTODO
Data care
Recordings remain under the Nation's authority. Any release of derived data requires the Language Department's approval, and restricted or ceremonial recordings are excluded outright.
07

Training

HyperparameterValue
Base modelopenai/whisper-small
MethodLoRA
Learning rateTODO
Batch sizeTODO
Epochs / stepsTODO
OptimizerTODO
HardwareTODO
bash
# TODO: confirm the exact training invocation
python train.py \
  --base-model openai/whisper-small \
  --method lora \
  --lora-rank TODO --lora-alpha TODO --lora-dropout TODO \
  --learning-rate TODO --batch-size TODO --epochs TODO \
  --dataset cpn-bodewadmimwen
08

Evaluation

MetricValue
Word error rate (WER)38.6%
Character error rate (CER)11.2%
Evaluation setTODO
Held-out speakersTODO

These numbers describe performance on held-out community audio and should be read as a usable-with-review system, not a finished one. At 38.6% WER, output is a strong first pass for a human transcriber and a poor substitute for one. The low CER means errors tend to be near-misses in morphology and spelling rather than lost content.

Note
No comparison baseline exists: there was no prior ASR system for Bodéwadmimwen to measure against.
09

Inference API

Request fieldDetail
EndpointTODO
MethodPOST (multipart/form-data)
Fieldaudio — WAV file
AuthTODOscheme to confirm
json
{
  "text": "…transcribed Bodéwadmimwen text…",
  "duration_s": 4.12
}
// TODO: confirm the exact response schema and error payloads
TODO
Error responses — status codes, payload shape, and rate limits — are not yet documented.
10

Model card

Intended use. Assisting transcription of Bodéwadmimwen audio in language-documentation and learning workflows, with a human reviewing output; powering pronunciation feedback in the Nishnabe Bawaajigan companion app.

Out of scope. Unreviewed publication of transcripts; legal, medical, or evidentiary transcription; speaker identification or surveillance of any kind; transcription of other Anishinaabe languages, which this model was not trained on.

Known limitations. Trained on under ten hours from a small speaker set, so accuracy degrades on unfamiliar voices, dialectal variation, noisy or far-field recordings, and speech outside the recorded domains. Orthographic conventions follow the training transcripts and may not match other writing systems for the language.

11

Data governance

This documents a community's language, so governance is a product requirement rather than a disclaimer. Work follows the CARE Principles for Indigenous Data Governance — Collective benefit, Authority to control, Responsibility, Ethics — alongside, not instead of, the technical FAIR framing.

TermStatus
Ownership of recordingsTODOto be documented with CPN
Ownership of trained modelTODO
License / permitted useTODO
Consent basis for training audioTODO
Review / restrict / revoke processTODO
Community authority
The Citizen Potawatomi Nation retains authority over the language data this system was built on, including the ability to review, restrict, or revoke its use. The governance and acknowledgment content on this site is reviewed by the Nation before public launch.
12

Companion app

Nishnabe Bawaajigan is the learning app built on this model, giving learners spoken practice with feedback rather than text drills alone. It is a separate product; this site documents only the recognition system underneath it.

DetailValue
PlatformTODO
LinkTODO
13

Citation

If you reference this system, cite it as below and acknowledge the Citizen Potawatomi Nation Language Department as the source of the language data.

bibtex
@misc{mootakoduru_bodewadmimwen_asr,
  author       = {Mootakoduru, Akshay},
  title        = {Bodéwadmimwen ASR: Speech Recognition for Potawatomi},
  howpublished = {Built with the Citizen Potawatomi Nation Language Department},
  year         = {TODO},
  url          = {TODO}
}
14

Acknowledgments

This work exists because of the Citizen Potawatomi Nation Language Department and the speakers whose voices are in the recordings. Particular thanks to Justin Neely, Director of the Language Department, for the partnership that made the corpus and this system possible.

The language belongs to the community, not to the model. Any credit for this system belongs first to the people who kept Bodéwadmimwen spoken long enough for it to be recorded.