Polyphase resampling (no aliasing); n_perm switch for tests

app.py

@@ -33,16 +33,20 @@ SR = 16000
 
 
 def _to16k(audio):
+    """Gradio gives (sr, int16 array); resample with a polyphase filter (plain interpolation aliases and makes
+    speech sound harsher, which skews emotion and speaking-style answers)."""
+    from math import gcd
+    from scipy.signal import resample_poly
     sr, x = audio
     x = np.asarray(x)
     if x.dtype.kind == "i":
-        x = x.astype(np.float32) / np.iinfo(x.dtype).max
+        x = x.astype(np.float32) / float(np.iinfo(x.dtype).max + 1)
     x = x.astype(np.float32)
     if x.ndim == 2:
         x = x.mean(axis=1)
     if sr != SR:
-        n = int(round(len(x) * SR / sr))
-        x = np.interp(np.linspace(0, len(x) - 1, n), np.arange(len(x)), x).astype(np.float32)
+        g = gcd(int(sr), SR)
+        x = resample_poly(x, SR // g, int(sr) // g).astype(np.float32)
     return x[: SR * 20]  # at most 20 s
 
 
@@ -71,20 +75,20 @@ def parse_table(text, lang):
 
 
 @spaces.GPU(duration=20, size=GPU_SIZE)
-def _run(x, qs):
+def _run(x, qs, n_perm):
     t = time.perf_counter()
-    res = DECIDER.decide(x, qs) if MOCK else DECIDER.decide(x, qs, n_perm=N_PERM)
+    res = DECIDER.decide(x, qs) if MOCK else DECIDER.decide(x, qs, n_perm=n_perm)
     return res, (time.perf_counter() - t) * 1000
 
 
-def decide(audio, lang, text):
+def decide(audio, lang, text, n_perm=None):
     if audio is None:
         raise gr.Error("Record or upload a clip first (or pick an example).")
     x = _to16k(audio)
     if len(x) < SR * 0.2:
         raise gr.Error("The clip is too short.")
     qs = parse_table(text, lang)
-    res, ms = _run(x, qs)
+    res, ms = _run(x, qs, int(n_perm or N_PERM))
     rows = []
     for q in qs:
         r = res[q.id]
@@ -135,6 +139,7 @@ with gr.Blocks(title="DuplexJev — speech to decisions", theme=gr.themes.Soft(p
             with gr.Accordion("Questions (editable) / 判断题（可编辑）", open=False):
                 table = gr.Textbox(value=table_text("en"), lines=11, show_label=False)
             btn = gr.Button("Decide ⚡", variant="primary")
+            nperm = gr.Number(value=N_PERM, visible=False)
             gr.Examples(EXAMPLES, inputs=[audio, lang], label="Examples (synthetic or public-licence clips)",
                         example_labels=["EN · angry barge-in", "EN · stops mid-request", "EN · “mm-hmm” backchannel", "laughter", "中文 · 等一下我想想", "中文 · 后排闲聊（不是对助手说）", "中文 · 开心道谢", "中文 · 导航请求"])
         with gr.Column(scale=6):
@@ -143,7 +148,7 @@ with gr.Blocks(title="DuplexJev — speech to decisions", theme=gr.themes.Soft(p
                                interactive=False, max_height=1000, column_widths=["18%", "34%", "13%", "35%"])
     gr.Markdown(FOOT)
     lang.change(set_lang, lang, table)
-    btn.click(decide, [audio, lang, table], [head, out], api_name="decide")
+    btn.click(decide, [audio, lang, table, nperm], [head, out], api_name="decide")
 
 if __name__ == "__main__":
     demo.queue(max_size=40).launch()