Inconsistent padding_bonus effect: TTS sometimes speaks fast, sometimes at default speed

[Isam-tfares]

Due diligence

• I have done my due diligence in trying to find the answer myself.

Topic

The paper

Question

Summary

I'm using the moshi TTS (Rust server + Python module) and trying to control the speech speed using the padding_bonus parameter in the config.toml. I’ve set:

padding_bonus = -2

I expected this to consistently make the TTS speak faster, which sometimes works. But in many cases, the TTS still speaks at normal/default speed as if the parameter is ignored.

---

My config.toml :

static_dir = "./static/"
log_dir = "/tmp/unmute_logs"
instance_name = "tts"
authorized_ids = ["public_token"]

[modules.tts_py]
type = "Py"
path = "/api/tts_streaming"
text_tokenizer_file = "hf://kyutai/tts-1.6b-en_fr/tokenizer_spm_8k_en_fr_audio.model"
# A higher batch size allows you to serve more users at once, but with a higher latency and memory usage.
batch_size = 4
text_bos_token = 1

[modules.tts_py.py]
log_folder = "/tmp/unmute_logs"
# We could use replace **/*.safetensors with unmute-prod-website/*.safetensors
# to only get the voices used in Unmute, but we are using the TTS for the demo
# on the project page too and for that we want to load the other voices as well
voice_folder = "hf-snapshot://kyutai/tts-voices/**/*.safetensors"
default_voice = "cml-tts/fr/10087_11650_000028-0002.wav"
cfg_is_no_text = true
n_q = 24
padding_bonus = -2
cfg_coef = 2.0

---

Behavior

• Sometimes: the voice is clearly faster (as expected).
• Other times (even during the same runtime or request type): the voice plays at the normal/default speed.
• I'm using moshi with the Rust server and calling the /api/tts_streaming endpoint.
• This seems non-deterministic. The parameter padding_bonus does not always apply.

---

Questions

1. Is there any known condition where padding_bonus is ignored (e.g., short input, voice model fallback, etc.)?
2. Are there internal voice/model parameters that override it?
3. Should I set this parameter differently (e.g., via API call instead of config.toml)?

---

Any clarification would be appreciated. I’d be happy to test or debug if needed.
Thanks!

[LaurentMazare]

I'm afraid that's the usual behavior of the padding_bonus, it's mostly a hack and shouldn't be relied upon too much.

[Isam-tfares]

Thanks for the clarification!
Is there another way to reliably control or force the speech speed?