usamaehsan/llm-ears
Research: gives a frozen LLM (Qwen3-8B) ears through a closed-form linear bridge from a frozen Whisper large-v3 encoder. No gradient training. Experiment mode measures WER on LibriSpeech.
Run usamaehsan/llm-ears with an API
Use one of our client libraries to get started quickly. Clicking on a library will take you to the Playground tab where you can tweak different inputs, see the results, and copy the corresponding code to use in your own project.
Input schema
The fields you can use to run this model with an API. If you don't give a value for a field its default value will be used.
| Field | Type | Default value | Description |
|---|---|---|---|
| mode |
None
|
experiment
|
None
|
| configs |
string
|
[{"layer": 32, "ctx": [0, 4, 12], "freq_pow": 0.5}, {"layer": 32, "ctx": [0, 4, 12], "freq_pow": 1.0}, {"layer": 32, "ctx": [0, 2, 4, 8, 12, 20], "freq_pow": 0.75}]
|
experiment: JSON list of {layer (1-32), ctx (list of context radii in frames), rff (random ReLU features, 0=off), freq_pow (inverse token-frequency frame weight, 0=off)}
|
| grid |
string
|
{"alpha": [0.001, 0.01, 0.03, 0.1, 0.3], "smooth": [3, 5, 7], "sil_bias": [0.0, 0.05, 0.1, 0.2], "min_run": [2, 3, 4, 6]}
|
experiment: JSON search grid tuned on a 5% holdout
|
| n_fit |
integer
|
0
|
experiment: max dev-clean utterances (0 = all 2703)
|
| n_train100 |
integer
|
0
Max: 28539 |
experiment: also fit on this many train-clean-100 utterances (downloaded on first use)
|
| n_test |
integer
|
400
Min: 10 Max: 2620 |
experiment: test-clean utterances for WER
|
| n_llm |
integer
|
200
Max: 2620 |
experiment: test utterances also read by the LLM
|
| alignments |
string
|
experiment: alignments.json.gz from an earlier run
|
|
| seed |
integer
|
0
|
None
|
| audio |
string
|
listen: audio file
|
|
| bridge |
string
|
listen: bridge.safetensors from an experiment run
|
|
| question |
string
|
Write exactly what was said.
|
listen: what to ask the LLM
|
{
"type": "object",
"title": "Input",
"properties": {
"grid": {
"type": "string",
"title": "Grid",
"default": "{\"alpha\": [0.001, 0.01, 0.03, 0.1, 0.3], \"smooth\": [3, 5, 7], \"sil_bias\": [0.0, 0.05, 0.1, 0.2], \"min_run\": [2, 3, 4, 6]}",
"x-order": 2,
"description": "experiment: JSON search grid tuned on a 5% holdout"
},
"mode": {
"enum": [
"experiment",
"listen"
],
"type": "string",
"title": "mode",
"description": "An enumeration.",
"default": "experiment",
"x-order": 0
},
"seed": {
"type": "integer",
"title": "Seed",
"default": 0,
"x-order": 8
},
"audio": {
"type": "string",
"title": "Audio",
"format": "uri",
"x-order": 9,
"nullable": true,
"description": "listen: audio file"
},
"n_fit": {
"type": "integer",
"title": "N Fit",
"default": 0,
"minimum": 0,
"x-order": 3,
"description": "experiment: max dev-clean utterances (0 = all 2703)"
},
"n_llm": {
"type": "integer",
"title": "N Llm",
"default": 200,
"maximum": 2620,
"minimum": 0,
"x-order": 6,
"description": "experiment: test utterances also read by the LLM"
},
"bridge": {
"type": "string",
"title": "Bridge",
"format": "uri",
"x-order": 10,
"nullable": true,
"description": "listen: bridge.safetensors from an experiment run"
},
"n_test": {
"type": "integer",
"title": "N Test",
"default": 400,
"maximum": 2620,
"minimum": 10,
"x-order": 5,
"description": "experiment: test-clean utterances for WER"
},
"configs": {
"type": "string",
"title": "Configs",
"default": "[{\"layer\": 32, \"ctx\": [0, 4, 12], \"freq_pow\": 0.5}, {\"layer\": 32, \"ctx\": [0, 4, 12], \"freq_pow\": 1.0}, {\"layer\": 32, \"ctx\": [0, 2, 4, 8, 12, 20], \"freq_pow\": 0.75}]",
"x-order": 1,
"description": "experiment: JSON list of {layer (1-32), ctx (list of context radii in frames), rff (random ReLU features, 0=off), freq_pow (inverse token-frequency frame weight, 0=off)}"
},
"question": {
"type": "string",
"title": "Question",
"default": "Write exactly what was said.",
"x-order": 11,
"description": "listen: what to ask the LLM"
},
"alignments": {
"type": "string",
"title": "Alignments",
"format": "uri",
"x-order": 7,
"nullable": true,
"description": "experiment: alignments.json.gz from an earlier run"
},
"n_train100": {
"type": "integer",
"title": "N Train100",
"default": 0,
"maximum": 28539,
"minimum": 0,
"x-order": 4,
"description": "experiment: also fit on this many train-clean-100 utterances (downloaded on first use)"
}
}
}
Output schema
The shape of the response you’ll get when you run this model with an API.
{
"type": "object",
"title": "Output",
"properties": {
"text": {
"type": "string",
"title": "Text",
"nullable": true
},
"bridge": {
"type": "string",
"title": "Bridge",
"format": "uri",
"nullable": true
},
"report": {
"type": "string",
"title": "Report",
"format": "uri",
"nullable": true
},
"alignments": {
"type": "string",
"title": "Alignments",
"format": "uri",
"nullable": true
},
"transcript": {
"type": "string",
"title": "Transcript",
"nullable": true
}
}
}