Supported Evaluation Task Types
March 27, 2026 ยท View on GitHub
This document clearly lists the evaluation tasks supported by the current version of Omni-Eval Kit (o_e_Kit).
Our framework can flexibly combine different input and output modalities. The following evaluation task pipelines have been implemented and verified:
1. Speech + Text Input / Text Output
This is the typical speech question answering (Audio QA) or speech instruction understanding task.
- Input:
- An audio clip (
.wav) - An associated text question or instruction (
questionfield in.jsonl)
- An audio clip (
- Output:
- Model-generated text answer (
answerfield in.jsonl)
- Model-generated text answer (
- Typical evaluation metrics:
- Word Error Rate (WER)
- Character Error Rate (CER)
- Keyword Matching
2. Speech + Text Input / Speech Output
This is a multi-modal dialogue task with speech synthesis capability.
- Input:
- An audio clip (
.wav) - An associated text question or instruction (
.jsonl)
- An audio clip (
- Output:
- Model-generated speech response (saved as
.wavfile)
- Model-generated speech response (saved as
- Typical evaluation metrics:
- Objective metrics: e.g., Mel-Cepstral Distortion (MCD) computed by acoustic models.
- Subjective metrics (model scoring): e.g., using MOSNet (Mean Opinion Score) to predict the naturalness and quality of generated speech.
3. Video + Text Input / Text Output
This is the typical video content understanding or video question answering (Video QA) task.
- Input:
- A video clip (
.mp4) - An associated text question (
.jsonl)
- A video clip (
- Output:
- Model-generated text answer (
.jsonl)
- Model-generated text answer (
- Typical evaluation metrics:
- Accuracy
- VQA Score
4. Video + Text + Speech Input / Text Output
This is the most comprehensive omni-modal understanding task, requiring the model to process all three mainstream modalities simultaneously.
- Input:
- A video clip (
.mp4) - An associated text question (
.jsonl) - An associated user voice instruction or follow-up question (
.wav)
- A video clip (
- Output:
- Model-generated unified text answer (
.jsonl)
- Model-generated unified text answer (
- Typical evaluation metrics:
- Accuracy
- Task-specific metrics related to multi-modal fusion understanding.
This list will be updated as the project evolves.