CustomVoice language conditioning bug confirmed in v0.8.0 — working fix applied (ref #87 / PR #98)

Summary

The CustomVoice language conditioning bug documented in #87 (Issue B) and addressed by PR #98 is confirmed to still exist in v0.8.0. The assistantPreambleKernel in v0.8.0 still only implements the 8-row (no-language) prefix layout, producing unintelligible audio output for the only officially supported Qwen3-TTS model family (Qwen3-TTS-12Hz-0.6B-CustomVoice).

We have applied the fix from PR #98 (adapted for v0.8.0) and confirmed it produces correct, intelligible speech validated by human listening on a Jetson Orin Nano.

Environment

• TensorRT-Edge-LLM: v0.8.0 (main as of 2026-06-15)
• Device: Jetson Orin Nano (SM87, JetPack 6.x)
• CUDA: 13.2
• TensorRT: 10.x
• Model: Qwen/Qwen3-TTS-12Hz-0.6B-CustomVoice
• Build flags: -DENABLE_CUTE_DSL=gemm -DCUTE_DSL_ARTIFACT_TAG=sm_87 -DCMAKE_CUDA_ARCHITECTURES=87

What fails

Running qwen3_tts_inference against the 0.6B-CustomVoice model with v0.8.0 unmodified produces WAV files that contain:

• Very short bursts of clicks/hiss (~1-2 seconds)
• Longer hum/noise with no speech content (~13 seconds when using user role in messages)

The audio is structurally valid (WAV container, correct sample rate) but semantically wrong — no intelligible speech.

Root cause (same as #87 Issue B)

assistantPreambleKernel in cpp/kernels/talkerMLPKernels/talkerMLPKernels.cu only implements the 8-row (Base / no-language) prefix:

rows 0-2: projected[0-2]
row 3:    ttsPad + embTable[codecNothinkId]
row 4:    ttsPad + embTable[codecThinkBosId]
row 5:    ttsPad + embTable[codecThinkEosId]
row 6:    ttsPad + embTable[speakerId]
row 7:    ttsBos + embTable[codecPadId]

CustomVoice models require the 9-row layout with language_id:

rows 0-2: projected[0-2]
row 3:    ttsPad + embTable[codecThinkId]
row 4:    ttsPad + embTable[codecThinkBosId]
row 5:    ttsPad + embTable[languageId]        ← NEW
row 6:    ttsPad + embTable[codecThinkEosId]
row 7:    ttsPad + embTable[speakerId]
row 8:    ttsBos + embTable[codecPadId]

Additionally, codec_language_id is not propagated from the HuggingFace talker_config during export (tensorrt_edgellm/scripts/export.py — _patch_tts_config copies codec_think_id but not codec_language_id).

Fix applied (v0.8.0)

We applied the equivalent of PR #98 adapted for v0.8.0. Files changed:

1. cpp/kernels/talkerMLPKernels/talkerMLPKernels.{cu,h} — Added languageId, codecThinkId, and prefixRows parameters. When languageId >= 0, emits the 9-row prefix; otherwise preserves the 8-row layout.

2. cpp/runtime/qwen3OmniTTSRuntime.{cpp,h} — Added mLanguageIdMap and mActiveLanguageId. Parses codec_language_id map and codec_think_id from config.json. Auto-selects "english" as default when available. Threads languageId through projectToTalkerInput and buildTalkerPrefillFromSegments.

3. examples/omni/qwen3_tts_inference.cpp — Forwards "language" field from input JSON.

4. tensorrt_edgellm/scripts/export.py — Added "codec_language_id" to the list of keys copied from talker_config in _patch_tts_config.

5. Engine config.json — Manually added the 12-language codec_language_id map from the HuggingFace checkpoint's talker_config.

Validation

After the fix:

projectToTalkerInput: seqLen=21, N=13, prefixRows=9, outputSeqLen=24, speakerId=3066, languageId=2050
Batch 0: 47 audio frames (exit: EOS)

• Output: 3.76s of clear, intelligible English speech ("Hello, I am Chip. It is nice to meet you.")
• Validated by human listening on Jetson Orin Nano via reSpeaker XVF3800 USB audio
• Speaker: serena (id=3066), Language: english (id=2050)

Additional note on message format

The official TTS docs show messages with "role": "assistant" only (not "user"). Using "role": "user" produces long silence/hum even with the language fix applied. This is consistent with Qwen3-TTS being a pure TTS model, not conversational.

Request

PR #98 addresses this exact issue but targets v0.7.x and currently has merge conflicts (dirty status). Could the maintainers either:

1. Rebase/merge PR #98, or
2. Apply the equivalent fix to v0.8.0 main

This is the only bug preventing the officially supported CustomVoice models from producing correct output.

References

• #87 — Original bug report (Issue B: language conditioning)
• #98 — Community PR with the fix (v0.7.x, merge-dirty)
• Qwen3-TTS CustomVoice config.json — source of codec_language_id map

Start by comparing PR #98 with the v0.8.0 versions of cpp/kernels/talkerMLPKernels/talkerMLPKernels.cu, cpp/runtime/qwen3OmniTTSRuntime.cpp, examples/omni/qwen3_tts_inference.cpp, and tensorrt_edgellm/scripts/export.py. Run qwen3_tts_inference with the CustomVoice model and verify the language-aware prefix produces clear speech and preserves the no-language path.