QWEN × FISHDACRESEARCH NOTE / 06 NOV 2026
APPROACH B · SINGLE-SPEAKER STUDY

A voice,
reconstructed.

One voice. Three listening views.
Listen to what survives the journey through a frozen Qwen codec and a fine-tuned FishDAC decoder.

Enter the listening room ↗
PHIL BENSON / SPEAKER 6097Loading verified snapshot…
CHECKPOINT—Completed updates
TRAINING CORPUS29.95 h39,373 clean utterances · one speaker
OUTPUT SAMPLE RATE44.1 kHzOriginal amplitude · float WAV
LISTENING EXAMPLES6616 training + all 50 validation clips
THE SNAPSHOT

Loading measured results…

↓

01 / LISTEN CLOSELY

The listening room

Matched utterances. No loudness normalization.
Switch signals at the same playback position.

HELD-OUT VALIDATION
SOURCE TRANSCRIPT
Loading…

Ground truth

A / ORIGINAL RECORDING↓ WAV
0:00 / 0:00

Reconstruction

B / MODEL OUTPUT↓ WAV
0:00 / 0:00

Reconstruction · 24 kHz

C / BAND-LIMITED CONTROL↓ WAV
0:00 / 0:00

What happens above 12 kHz?

The third player is the same reconstruction, anti-aliased and downsampled to 24 kHz—not another model. Compare all three mel spectrograms on identical time, frequency and color scales.

Ground truth, full-band reconstruction and 24-kHz reconstruction mel spectrograms, with shared scales and a dashed 12-kHz line
↓ Download spectrogramClick the image to inspect at full resolution ↗

Dashed line: 12-kHz Nyquist limit of the 24-kHz control. Color: fixed −100 to −20 dB; no per-panel normalization. High-band energy alone can be noise: fidelity requires agreement with the ground truth. The control is upsampled only for identical analysis axes; its audio download stays at 24 kHz.

Inspect the signal Waveforms, spectrum & source ↗

Welch power spectral density, 2,048-sample windows. Mint: ground truth; lavender: reconstruction; amber: 24-kHz control. All use the same axes; this is not a perceptual quality score.

Waveform envelopes share one peak scale across all three signals. No gain normalization or fitted time shift is applied.

Validation = official dev.clean, excluded from this fine-tuning run. Training examples are a fixed seeded panel, not the full training set. The 100 official test clips are not used here.

02 / FOLLOW THE FIT

A trajectory, not a verdict.

Unsmoothed, fixed-panel evaluation.
Lower is better for both reconstruction losses.

Train · fixed 16 Validation · all 50x = additional single-speaker updates

Mel-spectrogram loss

Multi-resolution STFT loss

Each point is a per-clip mean on the same 16 train / 50 validation utterances. These are evaluation losses—not minibatch training curves, MOS, or proof of perceptual equivalence. Use the audio to judge residual artifacts.

03 / WHAT YOU ARE HEARING

Frozen context.
Learned synthesis.

The full utterance passes through the frozen Qwen encoder and the retained beginning of its native decoder. A trainable bridge then conditions FishDAC’s synthesis stack.

This is a single-speaker fitting study, not an unseen-speaker benchmark. Validation shares the speaker and source books with training. Upstream pretraining overlap has not been audited.
Frozen Qwen path
Encoder + RVQ → decoder pre-convolution → full 8-block pre-transformer → first ×2 upsampler + ConvNeXt.
The bridge
1,024 features @ 25 Hz → fixed rate conversion to ≈21.53 Hz → trainable adapter.
Trainable FishDAC
Both latent ×2 upsamplers + ConvNeXt, then the waveform decoder → 44.1 kHz. Fish post-transformer bypassed.
Fine-tuning
Warm-started from B step 10,950, including G/D optimizer state. Global batch 8; constant LR: bridge 2e-4, synthesis 5e-5, discriminator 2e-5. No step cap.
Objective
Mel + MR-STFT + GAN + feature matching + frozen Qwen-ASR perceptual loss. Native gain, no data augmentation.
Evaluation
Fresh full-utterance inference from one immutable checkpoint. 198 float32 WAV files: original, full-band reconstruction and 24-kHz control. Fresh losses verified against saved evaluation.

OPEN THE EVIDENCE

Reproducible by design.

Data: original Hi-Fi TTS / OpenSLR 109, NVIDIA et al., CC BY 4.0; speaker 6097, clean partition. Ground-truth WAVs preserve the original samples. Reconstructions are experimental model outputs. Original chapter recordings were MP3-sourced; 44.1 kHz does not imply pristine full-band originals.

↓ Metrics, samples & hashes JSON↓ Method & reproduction MD
Snapshot identity & limitations

Speaker selection used recording-quality proxies and development audio. No MOS study or blind listening study is claimed. This is a static report snapshot; training may have advanced since export.