README.md

---
pretty_name: KambaBench-ASR
language:
  - kam
license: cc-by-4.0
task_categories:
  - automatic-speech-recognition
tags:
  - kamba
  - kikamba
  - kenya
  - low-resource
  - speech
  - asr
  - benchmark
  - african-languages
size_categories:
  - n<1K
---

# KambaBench-ASR

**Status: v0.0 — scaffold. No evaluation audio or gold transcriptions have been finalized yet.**

An open, leakage-controlled, reproducible evaluation benchmark for **Kamba (Kikamba, `kam`) automatic speech recognition (ASR)**.

KambaBench-ASR is designed to provide a common evaluation standard for Kamba speech-recognition systems. The benchmark is intended to be model-agnostic: any Kamba ASR system, whether based on Whisper, MMS, Omnilingual ASR, Parakeet, or another architecture, should be able to evaluate against the same frozen test set and scoring protocol.

The goal is simple:

> **Build a trustworthy, community-oriented benchmark for Kamba speech recognition that makes model performance measurable, comparable, and reproducible.**

This repository is the **benchmark, not a model**.

---

## Why KambaBench-ASR exists

Kamba already has access to a growing amount of speech and language data from multiple sources, including:

- DDD-Kenya's `Kamba-ASR-Data-Subset-484H`
- The Thiomi project
- Google FLEURS `kam_ke`
- Meta's Omnilingual ASR corpus

Together, these resources provide a promising foundation for developing and evaluating Kamba speech-recognition systems.

However, having speech data is not the same as having a reliable benchmark.

Before Kamba ASR systems can be compared fairly, several problems need to be addressed:

- **Data leakage:** the same recordings may appear across multiple datasets or may already have been used to train an evaluated model.
- **Different speech domains:** datasets may contain read speech, translated sentences, scripted recordings, or spontaneous speech.
- **Orthographic variation:** different valid or common Kamba spelling conventions can make WER appear worse even when the recognized speech is substantially correct.
- **Dialect variation:** Kamba speech may vary across speakers and regions.
- **Speaker overlap:** the same speaker may appear across training and evaluation datasets.
- **Transcript quality:** evaluation transcripts need to be checked and validated by native Kamba speakers.
- **Inconsistent evaluation:** different projects may use different preprocessing, normalization, decoding and scoring methods.

KambaBench-ASR aims to establish a single, transparent evaluation process that addresses these issues.

---

# Benchmark principles

KambaBench-ASR is built around five principles:

### 1. Leakage control

Evaluation audio must not be used to train the model being evaluated.

Candidate datasets will be checked for:

- speaker overlap
- filename collisions
- audio-hash collisions
- text/transcription overlap
- cross-corpus duplication

Where possible, the benchmark will document the provenance of every evaluation clip.

---

### 2. Native-speaker validation

A Kamba transcription should not be considered final simply because it exists in a dataset.

Evaluation clips should be reviewed by native Kamba speakers to confirm that:

- the audio is intelligible,
- the transcription accurately represents what was spoken,
- spelling is acceptable,
- dialectal forms are correctly represented,
- and problematic recordings are identified.

A clip will only contribute to the benchmark's **validated headline score** after native-speaker verification.

---

### 3. Kamba-aware normalization

Kamba ASR should not be evaluated solely through raw string matching.

KambaBench-ASR will investigate and document a language-aware normalization process that handles legitimate spelling and formatting variation without changing the underlying linguistic content.

The same normalization rules will be applied to every evaluated system.

The benchmark will report both:

- **raw scores**
- **normalized scores**

This makes it possible to distinguish genuine recognition errors from differences caused primarily by orthographic representation.

---

### 4. Multiple evaluation dimensions

A single WER number does not tell the whole story.

KambaBench-ASR will aim to report:

- Character Error Rate (CER)
- Word Error Rate (WER)
- normalized CER
- normalized WER
- dialect-level performance
- speaker-level performance where appropriate
- domain-level performance
- read vs. spontaneous speech performance

CER will be particularly important when evaluating orthographic variation.

---

### 5. Reproducibility

Every benchmark release should be reproducible.

The benchmark will maintain:

- frozen evaluation manifests
- dataset provenance
- deterministic scoring
- versioned normalization rules
- checksums/hashes where appropriate
- documented evaluation procedures
- pinned software dependencies

A model should be evaluated against the same benchmark version regardless of who runs the evaluation.

---

# Planned repository structure

```text
kambabench/
  tasks/
    asr/
      eval_unscripted.jsonl
      SCHEMA.md
  scoring/
    score_asr.py
    kb_io.py
    normalize.py
    dedup_check.py
    domain_audit.py
    requirements.txt
  docs/
    DATASETS.md
    EVALUATION.md
    ORTHOGRAPHY.md
    VALIDATION.md
    PROVENANCE.md
  reports/
    benchmark_report.md
  README.md
```

### `eval_unscripted.jsonl`

The frozen evaluation manifest will contain information such as:

```text
id
audio_pointer
gold_transcription
domain
dialect
speaker_id_hash
source_corpus
provenance
in_training_corpus
needs_native_validation
validation_status
```

The benchmark will avoid exposing personally identifying speaker information.

---

# Evaluation integrity

KambaBench-ASR will use an **evaluation integrity firewall**.

A clip can contribute to the validated headline score only when:

```text
native_validation = true
AND
in_training_corpus = false
AND
provenance_verified = true
```

If any of these conditions are not satisfied, the clip may still be useful for development or exploratory evaluation, but its result must not be presented as a validated benchmark score.

This distinction is important.

There is a difference between:

> **"The model achieved this score on the available evaluation set."**

and:

> **"The model achieved this validated benchmark score on a leakage-controlled, native-speaker-verified test set."**

KambaBench-ASR aims for the second.

---

# Candidate source corpora

The following resources are potential sources for training, validation, benchmarking, or cross-corpus analysis.

| Source | Reported scale | Primary use |
| --- | ---: | --- |
| **DDD-Kenya — Kamba-ASR-Data-Subset-484H** | ~484 hours / ~161k clips | Large-scale Kamba ASR training |
| **Thiomi** | ~54.5k approved Kamba recordings reported by the project; public `thiomi-5k` is a smaller release | Training, validation and natural-speech research |
| **Google FLEURS — `kam_ke`** | ~1,000 train / ~400 dev / ~400 test | Standardized evaluation and read speech |
| **Meta Omnilingual ASR corpus — `kam_Latn`** | ~51.1 hours reported | Multilingual/spontaneous speech and model evaluation |

These datasets should **not automatically be combined**.

Before using any source in the benchmark, its:

- license,
- provenance,
- speaker composition,
- recording conditions,
- transcription methodology,
- domain,
- dialect coverage,
- and relationship to other datasets

must be documented.

---

# Cross-corpus leakage audit

Before an evaluation clip is included in a final benchmark release, candidate datasets should be compared against one another.

The audit will check for:

### Speaker overlap

```text
same speaker → potential leakage
```

### Filename overlap

```text
same filename → investigate
```

### Audio-hash overlap

```text
same audio → definite duplicate
```

### Text overlap

```text
same transcript → investigate
```

Text overlap alone does not necessarily mean audio leakage. For example, two datasets may independently record the same publicly available sentence.

The benchmark will therefore distinguish between:

- text overlap,
- speaker overlap,
- audio duplication,
- and confirmed training-data leakage.

---

# Speech-domain coverage

KambaBench-ASR will aim to avoid creating a benchmark that only measures performance on carefully read sentences.

Where sufficient data is available, evaluation should include multiple speech conditions:

### Read speech

Controlled recordings of prepared text.

### Spontaneous speech

Natural conversation and unscripted responses.

### Domain-specific speech

Potential domains include:

- education
- agriculture
- healthcare
- finance
- government services
- customer service
- everyday conversation
- technology

### Noisy speech

Where suitable data is available, recordings with realistic background noise should also be considered.

The benchmark will document the composition of the evaluation set rather than hiding it behind a single aggregate number.

---

# Kamba orthography

A dedicated Kamba normalization layer will be developed only after examining actual transcripts from the candidate datasets and consulting native Kamba speakers.

The normalizer should:

- preserve meaningful Kamba characters,
- handle punctuation consistently,
- normalize obvious formatting differences,
- document accepted spelling variants,
- avoid changing linguistic meaning,
- and remain deterministic.

The normalization rules will be versioned.

For example:

```text
KambaBench-ASR normalization v0.1
KambaBench-ASR normalization v0.2
...
```

This prevents benchmark results from changing silently when the normalizer is modified.

---

# Metrics

The primary metrics will be:

### Character Error Rate

```text
CER
```

CER measures character-level edit distance between the model output and the reference transcription.

### Word Error Rate

```text
WER
```

WER will remain useful but should be interpreted carefully when orthographic variation affects word boundaries or spelling.

### Normalized CER

```text
CER after KambaBench normalization
```

### Normalized WER

```text
WER after KambaBench normalization
```

All models will use the same scoring implementation.

---

# Models that can be evaluated

KambaBench-ASR is model-agnostic.

Potential systems include:

### Meta MMS

Kamba is supported by Meta's Massively Multilingual Speech models.

### Meta Omnilingual ASR

Kamba (`kam_Latn`) is included in the Omnilingual ASR ecosystem.

### Whisper

Whisper can be fine-tuned for Kamba using the available Kamba speech corpora.

### Existing Kamba Whisper models

Existing community Kamba Whisper fine-tunes can be used as baseline systems.

### NVIDIA Parakeet

Parakeet-TDT can be adapted to Kamba through Kamba-specific tokenizer and decoder adaptation.

### Future Kamba-specific models

Any future architecture should be able to participate as long as it can produce a Kamba transcription from audio.

---

# Baseline experiments

The initial benchmark should establish several baselines rather than immediately optimizing one model.

A proposed baseline matrix:

| Model | Purpose |
| --- | --- |
| MMS | Existing Kamba-capable baseline |
| Omnilingual ASR | Modern multilingual baseline |
| Whisper | Fine-tuning baseline |
| Existing Kamba Whisper | Community baseline |
| Parakeet-TDT | New Kamba adaptation |
| Human transcription agreement | Upper-bound/reference analysis |

This makes it possible to answer:

> **How much does Kamba-specific fine-tuning actually improve ASR performance?**

rather than simply reporting the performance of one model.

---

# What KambaBench-ASR is trying to achieve

The long-term goal is not simply to produce another Kamba speech-recognition model.

The goal is to establish infrastructure that makes **Kamba speech technology measurable and reproducible**.

A successful benchmark should make it possible for future researchers and developers to answer questions such as:

- Which ASR model currently performs best on Kamba?
- How well does Kamba ASR work on spontaneous speech?
- How much does fine-tuning improve existing multilingual models?
- Which Kamba dialects are underrepresented?
- How much does orthographic variation affect WER?
- How does performance change with more training data?
- How well do models perform on noisy real-world recordings?
- Can smaller models achieve competitive accuracy for on-device applications?

---

# Roadmap

## v0.0 — Scaffold

- [x] Define project scope
- [x] Define benchmark principles
- [x] Identify candidate datasets
- [ ] Confirm dataset provenance
- [ ] Confirm licenses
- [ ] Build ingestion pipeline
- [ ] Build cross-corpus deduplication tools

## v0.1 — Data audit

- [ ] Ingest candidate datasets
- [ ] Analyze speaker distribution
- [ ] Analyze domains
- [ ] Analyze dialect coverage
- [ ] Check cross-corpus overlap
- [ ] Identify candidate evaluation clips
- [ ] Establish provenance records

## v0.2 — Native validation

- [ ] Recruit native Kamba reviewers
- [ ] Validate evaluation transcripts
- [ ] Flag problematic audio
- [ ] Confirm dialect information
- [ ] Freeze evaluation set

## v0.3 — Scoring benchmark

- [ ] Implement CER
- [ ] Implement WER
- [ ] Implement Kamba normalization
- [ ] Add normalized metrics
- [ ] Freeze scoring protocol
- [ ] Publish benchmark report

## v1.0 — Public benchmark

- [ ] Release validated evaluation manifest
- [ ] Release scoring toolkit
- [ ] Publish baseline model results
- [ ] Publish benchmark documentation
- [ ] Assign a permanent benchmark version
- [ ] Invite community submissions

---

# Future work

Once the benchmark is established, Kamba ASR research can move beyond basic transcription accuracy.

Potential directions include:

- Kamba voice assistants
- Kamba speech-to-text applications
- Kamba-English translation
- Kamba-Swahili translation
- Kamba text-to-speech
- educational speech applications
- accessibility tools
- healthcare voice interfaces
- customer-service transcription
- on-device Kamba ASR
- code-switching detection
- dialect-aware speech recognition

---

# Contribution

KambaBench-ASR is intended to be community-driven.

Useful contributions include:

- Kamba speech data
- transcription review
- native-speaker validation
- orthography research
- dialect expertise
- benchmark tooling
- ASR models
- evaluation results
- documentation
- reproducibility testing

If you are a native Kamba speaker, linguist, researcher, developer, or organization working with Kamba language technology, your contribution can help improve the benchmark.

---

# Citation

If you use KambaBench-ASR, please cite the benchmark release and the underlying datasets used in your experiment.

Dataset-specific citations should be retained according to the requirements of each source corpus.

---

# License

The benchmark code, schema and documentation will be released under an open license.

Audio and transcription data will retain the licensing and usage restrictions of their original source datasets.

Where redistribution of source audio is not permitted, KambaBench-ASR will provide the appropriate metadata, provenance information, identifiers, and evaluation instructions rather than redistributing the underlying audio.

---

# Disclaimer

KambaBench-ASR is currently a **research scaffold**.

No benchmark score should be considered an official Kamba ASR result until the corresponding evaluation data has undergone the documented leakage, provenance and native-speaker validation process.

The benchmark prioritizes **transparency and reproducibility over headline numbers**.