Qwen3TTSBatchSession.step does not stream incrementally or emit final chunk on 0.4.3

[terrazorc]

Qwen3TTSBatchSession.step does not stream incrementally or emit final chunk on mlx-audio 0.4.3

Summary

Qwen3TTSBatchSession appears to be the canonical API for a fixed TTS model
serving many texts, but in mlx-audio==0.4.3 I cannot get it to behave like a
streaming API for Qwen3-TTS. With stream=True and streaming_interval=0.32,
session.step() emits one large audio event only after the utterance is already
substantially complete, then returns empty events until my safety cap is hit.
ev.is_final_chunk never becomes true.

This makes it hard to use the batch-session path for low-latency voice-agent TTS,
even though it looks like the right API shape for a persistent process with one
loaded voice/model and many short utterances.

Environment

• macOS: macOS-15.6-arm64-arm-64bit-Mach-O
• Machine: Apple Silicon / arm64
• Python: 3.14.4
• mlx-audio: 0.4.3
• mlx: 0.31.2
• mlx-metal: 0.31.2
• mlx-lm: 0.31.3
• Model tested: mlx-community/Qwen3-TTS-12Hz-0.6B-Base-8bit

Minimal reproducer shape

import time
from mlx_audio.tts.utils import load_model
from mlx_audio.tts.models.qwen3_tts.continuous_batching import (
    TTSBatchOptions,
    TTSBatchItem,
)

model = load_model("mlx-community/Qwen3-TTS-12Hz-0.6B-Base-8bit")
options = TTSBatchOptions(
    temperature=0.9,
    top_p=1.0,
    top_k=50,
    repetition_penalty=1.05,
    max_tokens=4096,
    lang_code="auto",
    stream=True,
    streaming_interval=0.32,
    max_batch_size=8,
    verbose=False,
)
session = model.create_tts_batch_session(options)

item = TTSBatchItem(
    sequence_id=0,
    text="First halo utterance to warm the session.",
    ref_audio="path/to/ref.wav",
    ref_text="reference transcript text",
)

t_submit = time.monotonic()
session.add([item])

for step_n in range(600):
    events = session.step()
    for ev in events:
        if ev.audio is not None:
            print(
                "first audio",
                round((time.monotonic() - t_submit) * 1000),
                "ms",
                "samples",
                ev.samples,
            )
        if ev.is_final_chunk:
            print("final", step_n)
            raise SystemExit(0)

print("hit safety cap without is_final_chunk")

Observed behavior

Three short texts in one session produced:

Text	First audio	Samples in first event	Completion
First halo utterance to warm the session.	2536 ms	71040	hit 600-step safety cap; no final chunk
Second utterance with more content here.	2339 ms	72960	hit 600-step safety cap; no final chunk
Third probe text.	1190 ms	38400	hit 600-step safety cap; no final chunk

After the first audio event, repeated session.step() calls returned empty
results. I did not observe incremental audio chunks matching
streaming_interval=0.32, and I did not observe is_final_chunk=True.

Expected behavior

With stream=True and streaming_interval=0.32, I expected one of these:

• multiple smaller audio events roughly as chunks become available, followed by
  a final event, or
• one audio event plus a reliable completion/final signal soon after.

Why this matters

For voice-agent TTS, the batch-session API is attractive because it should avoid
recreating per-request model/session state. In the current observed behavior, the
batch-session path has worse first-audio latency than our existing direct
model.generate(..., stream=True) path and has no reliable completion signal.

Is this expected for Qwen3-TTS in 0.4.3, or is there another required option /
calling pattern to make Qwen3TTSBatchSession produce incremental streaming
chunks and final events?