feat: Add more models

Author: dmtrKovalenko

src/screens/editor/Editor.res

@@ -209,7 +209,6 @@ let make = React.memo((
       <div className="flex justify-center w-full flex-1 min-h-0 overflow-hidden">
         {layout.mediaControls
         ->Belt.Option.map(size => {
-          Js.Console.log(size)
           let showSideBySide = size.width >= 770.0
 
           <div

src/screens/editor/Editor.res.mjs

@@ -29,12 +29,14 @@ export const MODELS = {
   "onnx-community/whisper-tiny_timestamped": "Tiny (152 Mb)",
   "onnx-community/whisper-base_timestamped": "Base (291 Mb)",
   "onnx-community/whisper-small_timestamped": "Small (586 Mb)",
+  "onnx-community/whisper-medium_timestamped": "Medium (769 Mb)",
+  "onnx-community/whisper-large-v3-turbo_timestamped": "Turbo (561 Mb)",
 } as const;
 
 export type Model = keyof typeof MODELS;
 
 export const ALL_MODELS = Object.keys(MODELS) as Array<Model>;
-const DEFAULT_MODEL: Model = "onnx-community/whisper-base_timestamped";
+const DEFAULT_MODEL: Model = "onnx-community/whisper-large-v3-turbo_timestamped";
 
 export const modelSerde: serde<Model> = {
   parse: function(input: string): Model {

src/transcriber/Constants.ts

@@ -20,6 +20,30 @@ async function isWebGPUAvailable() {
 // env.localModelPath = "/models/";
 // env.allowRemoteModels = true;
 
+// Large models that need aggressive quantization to fit in browser memory
+const LARGE_MODELS = [
+  "onnx-community/whisper-large-v3-turbo",
+  "onnx-community/whisper-large-v3-turbo_timestamped",
+  "distil-whisper/distil-large-v3",
+];
+
+// Get optimal dtype config based on model size
+function getDtypeConfig(modelName) {
+  if (LARGE_MODELS.some(m => modelName.includes(m.replace("onnx-community/", "").replace("distil-whisper/", "")))) {
+    // Use q4f16 for large models to fit in browser memory
+    // q4f16 = 4-bit quantized weights with fp16 activations
+    return {
+      encoder_model: "q4f16",
+      decoder_model_merged: "q4f16",
+    };
+  }
+  // Default config for smaller models
+  return {
+    encoder_model: "fp32",
+    decoder_model_merged: "q4", // or 'fp32' ('fp16' is broken)
+  };
+}
+
 // Define model factories
 // Ensures only one model is created of each type
 class PipelineFactory {
@@ -39,10 +63,7 @@ class PipelineFactory {
       this.instance = pipeline(this.task, this.model, {
         quantized: this.quantized,
         progress_callback,
-        dtype: {
-          encoder_model: "fp32",
-          decoder_model_merged: "q4", // or 'fp32' ('fp16' is broken)
-        },
+        dtype: getDtypeConfig(this.model),
         device: await isWebGPUAvailable() ? "webgpu" : "wasm",
       });
     }