Towards Zero-Shot Amplifier Modeling: One-to-Many Amplifier Modeling via Tone Embedding Control

---

Nine amplifiers performance of each model in Table 1

Comparison of the performance of different conditioning methods: Concat and FiLM with both Tone Embedding from paired reference and Look-Up Table (LUT) mechanisms, including a variant of using unpaired audio reference. The following audio samples present the results discussed in Section 4.2.

• Input : A clean audio extract from the test set.
• Ground truth : The corresponding amplifier-rendered audio extract from the test set.
• Concat + LUT : Audio generated using concatenation conditioning method and LUT embedding representation.
• Concat + Paired tone embedding : Audio generated using concatenation conditioning method and tone embedding extracted from paired reference audio.
• FiLM + LUT : Audio generated using FiLM conditioning method and LUT embedding representation.
• FiLM + Paired tone embedding : Audio generated using FiLM conditioning method and tone embedding extracted from paired reference audio.
• FiLM + Unpaired tone embedding : Audio generated using FiLM conditioning method and tone embedding extracted from unpaired reference audio.

	Input	Ground Truth	Concat + LUT	Concat + Paired tone embedding	FiLM + LUT	FiLM + Paired tone embedding	FiLM + Unpaired tone embedding
Amp 1
Amp 2
Amp 3
Amp 4
Amp 5
Amp 6
Amp 7
Amp 8
Amp 9

---

Zero shot on two additional amplifiers

We evaluate three reference tone embedding methods based on FiLM conditioning and unpaired reference. The following audio samples present the results discussed in Section 4.3.

• Input : A clean audio extract from the test set.
• Ground Truth : The corresponding unseen amplifier-rendered audio extract from the test set.
• Non-retrieval embedding : Same as the FiLM + Unpaired tone embedding setting, which does not change the reference tone embedding.
• Mean embedding : Replaces the reference tone embedding with an amplifier embedding that has the highest cosine similarity.
• Nearest embedding : Replaces the reference tone embedding with a seen tone embedding during training that has the highest cosine similarity.

	Input	Ground Truth	Non-retrieval embedding	Mean embedding	Nearest embedding
Amp 1 Low Gain
Amp 1 Low Gain
Amp 2 High Gain
Amp 2 High Gain

---

Zero shot on self-recorded audio

We test our method’s capability to model a self-recorded guitar audio with unseen content and tone from the training process.

• Lead L : Left channel of lead track.
• Lead R : Right channel of lead track.
• RH L : Left channel of rhythm track.
• RH R : Right channel of rhythm track.
• Solo : Solo track. Part of this clip is discussed and analysis in the section 4.4.

	Lead L	Lead R	RH L	RH R	Solo
Clean tone
Amplifier Rendered audio
audio generated from our method

We mixed these audio samples with bass, drum and keyboard tracks to test the overall quality in a real-world usage scenario. We recommend turning down the volume before listening to these results, as the dynamic range has not been compressed to preserve the characteristics of each track.