feat: limit transcription output length based on input

Using heuristics. Also adds documentation and initial unit tests.

ref: N25B-209

test/unit/agents/transcription/test_speech_recognizer.py

@@ -0,0 +1,36 @@
+import numpy as np
+
+from control_backend.agents.transcription import SpeechRecognizer
+from control_backend.agents.transcription.speech_recognizer import OpenAIWhisperSpeechRecognizer
+
+
+def test_estimate_max_tokens():
+    """Inputting one minute of audio, assuming 300 words per minute, expecting 400 tokens."""
+    audio = np.empty(shape=(60*16_000), dtype=np.float32)
+
+    actual = SpeechRecognizer._estimate_max_tokens(audio)
+
+    assert actual == 400
+    assert isinstance(actual, int)
+
+
+def test_get_decode_options():
+    """Check whether the right decode options are given under different scenarios."""
+    audio = np.empty(shape=(60*16_000), dtype=np.float32)
+
+    # With the defaults, it should limit output length based on input size
+    recognizer = OpenAIWhisperSpeechRecognizer()
+    options = recognizer._get_decode_options(audio)
+
+    assert "sample_len" in options
+    assert isinstance(options["sample_len"], int)
+
+    # When explicitly enabled, it should limit output length based on input size
+    recognizer = OpenAIWhisperSpeechRecognizer(limit_output_length=True)
+    options = recognizer._get_decode_options(audio)
+
+    assert "sample_len" in options
+    assert isinstance(options["sample_len"], int)
+
+    # When disabled, it should not limit output length based on input size
+    assert "sample_rate" not in options