whisper: set no_context to prevent quality drift over a session (#79)

* whisper: set no_context to prevent quality drift over a session

Whisper transcription quality degrades progressively over a long
push-to-talk session: short clips get mis-recognized or returned
empty, and language detection sticks to the previous language
(e.g. RU→EN switches keep producing Russian). Reloading the model
restores quality.

The cause is whisper.cpp's default prompt_past behaviour — the last
decoded tokens are fed back as a prompt for the next decode. That's
the right thing for continuous speech (lectures, meetings) where
consecutive segments are connected, but the wrong thing for
push-to-talk and similar workloads where each call to transcribe
is an independent utterance: stale prompt tokens bias the next
decode. Short clips suffer most because they have less acoustic
evidence to overcome the stale prompt; language switches suffer
because the prompt is in the previous language and steers detection.

Set no_context = true so each decode starts from a clean prompt.
The user-supplied initial_prompt continues to work — it goes through
a different field and is unaffected.

* whisper: expose no_context as a configurable field

Per review, make no_context an opt-in field on WhisperInferenceParams
so callers can override it for continuous-speech use cases (lectures,
meetings, streaming) where carrying prompt_past across segments
improves consistency. Default stays true — the right choice for
independent utterances such as push-to-talk dictation, which is the
case the previous commit fixed.

* fmt

---------

Co-authored-by: CJ Pais <cj@cjpais.com>

Author: anton-averich

src/onnx/session.rs

@@ -8,11 +8,11 @@ use ort::ep::ROCm;
 use ort::ep::TensorRT;
 #[cfg(feature = "ort-webgpu")]
 use ort::ep::WebGPU;
-#[cfg(feature = "ort-xnnpack")]
-use ort::ep::XNNPACK;
 use ort::ep::CPU;
 #[cfg(feature = "ort-cuda")]
 use ort::ep::CUDA;
+#[cfg(feature = "ort-xnnpack")]
+use ort::ep::XNNPACK;
 
 use ort::session::builder::GraphOptimizationLevel;
 use ort::session::Session;

src/whisper_cpp/mod.rs

@@ -110,6 +110,10 @@ pub struct WhisperInferenceParams {
 
     /// Initial prompt to provide context to the model.
     pub initial_prompt: Option<String>,
+
+    /// Start each decode with a clean prompt (whisper.cpp's `prompt_past`).
+    /// Default `true` suits push-to-talk; set `false` for continuous speech.
+    pub no_context: bool,
 }
 
 impl Default for WhisperInferenceParams {
@@ -126,6 +130,7 @@ impl Default for WhisperInferenceParams {
             no_speech_thold: 0.2,
             n_threads: 0,
             initial_prompt: None,
+            no_context: true,
         }
     }
 }
@@ -220,6 +225,7 @@ impl WhisperEngine {
         full_params.set_suppress_blank(params.suppress_blank);
         full_params.set_suppress_nst(params.suppress_non_speech_tokens);
         full_params.set_no_speech_thold(params.no_speech_thold);
+        full_params.set_no_context(params.no_context);
         if params.n_threads > 0 {
             full_params.set_n_threads(params.n_threads);
         }