diff --git a/src/Whisper.jl b/src/Whisper.jl
index 9183b7f..9f64c4f 100644
--- a/src/Whisper.jl
+++ b/src/Whisper.jl
@@ -14,6 +14,26 @@ function __init__()
     register_datadeps()
 end
 
+function create_ctx(model)
+    whisper_init_from_file(DataDeps.resolve("whisper-ggml-$model/ggml-$model.bin", "__FILE__"))
+end
+
+struct WhisperParams
+    ref::Base.RefValue{whisper_full_params}
+end
+Base.unsafe_convert(::Type{Ptr{whisper_full_params}}, wparams::WhisperParams) =
+    Base.unsafe_convert(Ptr{whisper_full_params}, getfield(wparams, :ref))
+Base.convert(::Type{whisper_full_params}, wparams::WhisperParams) =
+    getfield(wparams, :ref)[]
+Base.getproperty(wparams::WhisperParams, field::Symbol) =
+    getproperty(Base.unsafe_convert(Ptr{whisper_full_params}, wparams), field)
+Base.setproperty!(wparams::WhisperParams, field::Symbol, value) =
+    setproperty!(Base.unsafe_convert(Ptr{whisper_full_params}, wparams), field, value)
+function create_params(flag=LibWhisper.WHISPER_SAMPLING_GREEDY)
+    raw_params = whisper_full_default_params(flag)
+    return WhisperParams(Ref(raw_params))
+end
+
 """
     transcribe(model, data) -> String
 
@@ -24,9 +44,17 @@ automatically downloaded from HuggingFace on first use.
 - `data`: `Vector{Float32}` containing 16kHz sampled audio
 """
 function transcribe(model, data)
-    ctx = whisper_init_from_file(DataDeps.resolve("whisper-ggml-$model/ggml-$model.bin", "__FILE__") )
-    wparams = whisper_full_default_params(LibWhisper.WHISPER_SAMPLING_GREEDY)
+    ctx = create_ctx(model)
+    wparams = create_params(flag)
+
+    ret = transcribe(ctx, wparams, data)
 
+    whisper_free(ctx)
+
+    return ret
+end
+
+function transcribe(ctx, wparams, data)
     ret = whisper_full_parallel(ctx, wparams, data, length(data), 1)
 
     if ret != 0
@@ -44,8 +72,6 @@ function transcribe(model, data)
         @debug "Time for inference: ", t0-t1
     end
 
-    whisper_free(ctx)
-
     return result
 end