Avoid broadcast speaker injection (injecting embeddings to ALL codec positions)
in multi-speaker TTS training. Use when: (1) Training codec-based TTS models
with multiple speakers, (2) Model generates to max_new_tokens instead of
stopping at EOS, (3) Audio duration is excessively long (e.g., 163 seconds
for short text). Fix: Use single-position injection (position-6) which
preserves EOS detection. More speaker conditioning is NOT always better.