Update custom model files, README, and requirements

asr_pipeline.py

@@ -29,7 +29,7 @@ __all__ = [
     "SpeakerDiarizer",
 ]
 
-# Timestamps and speakers transcribe in chunks cut at the quietest point between
+# Audio is transcribed in chunks cut at the quietest point between
 # these lengths. The model trained on clips of at most 19 s; 18 leaves room for
 # the inference lead-in. Short clips are one chunk, so their text is unchanged.
 CHUNK_MAX_S = 18.0
@@ -144,7 +144,7 @@ class ASRPipeline(transformers.AutomaticSpeechRecognitionPipeline):
 
         try:
             if not return_timestamps:
-                return super().__call__(inputs, **kwargs)
+                return self._transcribe_plain(inputs, **kwargs)
             return self._transcribe_timed(
                 inputs,
                 return_speakers=return_speakers,
@@ -155,6 +155,36 @@ class ASRPipeline(transformers.AutomaticSpeechRecognitionPipeline):
             if original_prompt is not None:
                 self.model.TRANSCRIBE_PROMPT = original_prompt
 
+    def _transcribe_plain(self, inputs, **kwargs):
+        """Transcribe in the same chunks as `_transcribe_timed`, without timing words.
+
+        One call over a whole recording returns only its first
+        `max_new_tokens` of text (WER 97% on a 30-min AMI meeting), so plain
+        transcription chunks too. A short clip is one chunk: one call, the
+        same result as before. Inputs `_extract_audio` can't read (a batch)
+        go to the parent untouched.
+        """
+        audio = self._extract_audio(inputs)
+        if audio is None:
+            return super().__call__(inputs, **kwargs)
+        array = np.asarray(audio["array"], dtype=np.float32)
+        sr = audio.get("sampling_rate", 16000)
+
+        bounds = chunk_bounds(array, sr)
+        if len(bounds) == 1:
+            return super().__call__({"raw": array, "sampling_rate": sr}, **kwargs)
+        result: dict[str, Any] = {}
+        texts = []
+        for s, e in bounds:
+            out = super().__call__({"raw": array[s:e], "sampling_rate": sr}, **kwargs)
+            texts.append(out["text"])
+            # Per-step logprobs (output_scores=True) run on across chunks.
+            for key in ("top1_logprob", "top2_logprob"):
+                if key in out:
+                    result.setdefault(key, []).extend(out[key])
+        result["text"] = " ".join(t for t in texts if t)
+        return result
+
     def _transcribe_timed(
         self,
         inputs,