---

Abstract #

The music industry faces a persistent and costly challenge: determining whether a track will achieve viral reach before it is released to the public. Conventional approaches to music popularity prediction rely on post-publication engagement signals — streams, likes, shares, and historical interaction data — making them structurally incapable of informing pre-release decisions. GROMUS addresses this gap by introducing a unified five-network AI architecture capable of assessing viral potential from the raw audio signal alone, prior to any audience exposure. The system integrates five specialized neural networks operating on complementary signal modalities: a custom trend probability classifier that assigns High, Medium, or Low virality class; an automatic viral segment detector that identifies the most impactful audio fragment; a vibe and genre embedding module that measures stylistic proximity to trending sounds; a lyrics trend analyzer powered by OpenAI Whisper transcription; and a unified decision core that aggregates all sub-scores into a single, explainable FinalTrendScore. Developed under institutional support from Odessa National Polytechnic University (Department of Economic Cybernetics) and deployed in production at gromus.ai, the architecture demonstrates that virality is a measurable intrinsic property of audio content — not merely an emergent social phenomenon. This paper formalizes the full mathematical model, benchmarks GROMUS against state-of-the-art post-publication methods, defines the evaluation framework, and establishes cross-domain relevance through parallel application in medical AI signal analysis.

1. Introduction #

The global music streaming economy generates hundreds of millions of track releases annually, yet only a small fraction achieve viral spread on short-form video platforms such as TikTok, Instagram Reels, and YouTube Shorts. The difference between a viral hit and an ignored release is frequently determined not by marketing budget or label support, but by measurable properties embedded in the audio itself — qualities of tempo, emotional texture, lyrical memorability, and sonic alignment with current trend aesthetics.

Despite this intuition, the dominant paradigm in music trend forecasting has remained reactive. Existing systems — whether based on autoregressive time series models, deep recurrent networks, or recommendation engines — operate on historical engagement data. They answer the question “how is this song currently performing?” rather than “will this song perform?” The distinction is critical: by the time a post-publication model can generate a meaningful forecast, significant marketing investment has already been made or foregone.

GROMUS (General Relevance and Output Music Unified System) was developed to invert this paradigm. The core thesis is that virality potential is a property of the sound itself — detectable through audio signal analysis, style embedding, and linguistic pattern recognition — and that a properly architected multi-network system can quantify it before a single listener has heard the track.

The architecture was developed by Oleh Ivchenko and Dmytro Hrybeniuk under the institutional framework of Odessa National Polytechnic University, Department of Economic Cybernetics. It is actively deployed in the Gromus AI platform at gromus.ai, where it serves music producers, independent artists, and label A&R teams seeking objective pre-release quality signals.

Research Questions #

RQ1: Can music virality be predicted before publication using audio signal analysis alone, without historical engagement data?

RQ2: How does a unified five-network architecture compare to single-model approaches for virality prediction?

RQ3: What role do lyrics and vibe similarity play as predictive features for social media music trends?

---

2. Existing Approaches (2026 State of the Art) #

2.1 Time Series Models for Post-Publication Popularity #

The most widely deployed methods for music trend analysis are time series models applied to streaming counts, chart positions, and social engagement metrics after release. ARIMA (AutoRegressive Integrated Moving Average) and its seasonal variant SARIMA have been applied to Spotify and Billboard chart data to model and forecast popularity trajectories (Li et al., 2021, arXiv:2110.15790). Facebook’s Prophet model, designed for business time series forecasting with seasonality decomposition, has been adapted for music streaming trend detection with moderate success.

Long Short-Term Memory (LSTM) networks represent the deep learning counterpart, capturing complex temporal dependencies in streaming sequences. Li et al. (2021) introduced LSTM-RPA, demonstrating improvements of 13–18% in F-score over standard LSTM baselines for music popularity prediction (DOI: 10.48550/arXiv.2110.15790). Recurrent architectures have shown superior performance over ARIMA on long-horizon prediction tasks where engagement patterns exhibit non-linear dynamics.

Fundamental limitation: All time series approaches — ARIMA, LSTM, Prophet — share the same structural constraint: they require post-publication engagement data. Without streams, views, or interaction history, they produce no meaningful output. This makes them entirely unsuitable for pre-release evaluation, which is precisely the moment when actionable decisions about mastering, promotion, and release timing must be made.

2.2 Audio Fingerprinting Systems #

Audio fingerprinting technologies, exemplified by Shazam’s landmark acoustic fingerprinting system, extract compact hash representations of audio content to enable exact or near-exact matching against a reference database. While highly effective for song identification, fingerprinting is fundamentally a retrieval task — it answers “what is this?” not “how will this perform?”

Su et al. (2022) identified danceability, energy, and tempo as key audio features correlated with virality on TikTok and Spotify (DOI: 10.3390/electronics11213518), but their approach remains descriptive rather than predictive in the pre-publication sense: it characterizes already-successful tracks rather than screening candidates before exposure.

2.3 Music Recommendation and Collaborative Filtering #

Recommendation systems based on collaborative filtering (user-item interaction matrices) and content-based filtering (audio feature similarity) represent another major paradigm. However, recommendation systems optimize for listener retention given known preferences — they require user history data, are inherently post-publication in design, and optimize for a different objective than viral reach prediction.

2.4 Audio Feature Extraction for Prediction #

The Music Information Retrieval (MIR) community has developed extensive methods for extracting acoustic features from raw audio (Lerch, 2014; DOI: 10.1007/978-3-319-09912-5_16). Mel-Frequency Cepstral Coefficients (MFCCs), spectral centroid, chroma features, zero-crossing rate, and rhythmic onset detection provide rich representations of timbral, harmonic, and rhythmic content. These features have been used in classification tasks predicting genre, mood, and chart success with varying accuracy.

The critical gap: existing MIR-based classifiers operate as single models predicting a single label from a flat feature vector. They do not decompose the prediction task into specialized sub-problems (segment detection, vibe similarity, lyrical analysis), do not provide explainability, and are not designed for the multi-dimensional nature of virality on short-form platforms where the relevant audio unit is a 15–30 second clip, not the full track.

2.5 Comparative Landscape #

The following diagram illustrates the capability boundaries of existing approaches relative to GROMUS:

graph TD
    A[Music Trend Prediction Methods] --> B[Time Series: ARIMA / LSTM / Prophet]
    A --> C[Audio Fingerprinting: Shazam]
    A --> D[Recommendation Systems]
    A --> E[Single-Model MIR Classifiers]
    A --> F[GROMUS Unified Architecture]

    B --> B1["✗ Requires post-publication data
✗ No audio analysis
✓ Good at trend tracking"]
    C --> C1["✗ Exact match only
✗ No virality signal
✓ Fast identification"]
    D --> D1["✗ Requires user history
✗ Not pre-publication
✓ Personalized recommendations"]
    E --> E1["△ Some pre-pub potential
✗ Single model, no decomposition
✗ Limited explainability"]
    F --> F1["✓ Pre-publication prediction
✓ Full audio analysis
✓ Segment detection
✓ Vibe + lyric signals
✓ Explainable multi-score output"]

    style F fill:#2d6a2d,color:#fff
    style F1 fill:#2d6a2d,color:#fff

---

3. Quality Metrics & Evaluation Framework #

Evaluating a pre-publication virality prediction system requires metrics distinct from standard music recommendation benchmarks. GROMUS defines four evaluation dimensions:

3.1 Virality Classification Metrics #

For the three-class classification output (High / Medium / Low virality):

Metric	Definition	Target
Macro F1-Score	Harmonic mean of precision and recall across all three classes	≥ 0.78
High-class Precision	True viral tracks / All predicted viral	≥ 0.75
High-class Recall	True viral tracks correctly identified / All true viral	≥ 0.80
Low-class Specificity	True non-viral / All predicted non-viral	≥ 0.85

3.2 Segment Detection Accuracy #

The viral segment detector must localize the most-used audio fragment — the clip that users actually select for their videos. Evaluation uses:

• Intersection over Union (IoU): Overlap between predicted segment boundaries and human-labeled “high-engagement” windows from post-publication data
• Segment Rank Accuracy: Whether the top-ranked predicted segment appears in the top-3 actually used segments (target: ≥ 72%)

3.3 Vibe Similarity Scoring #

The vibe embedding module is evaluated by:

• Cosine Similarity Calibration: Correlation between predicted VibeScore and post-publication engagement on trend-aligned content
• Genre Cluster Purity: Whether embeddings cluster correctly by trend category (measured by silhouette coefficient)

3.4 Aggregated Evaluation Framework #

flowchart LR
    A[Ground Truth Labels\nPost-publication virality] --> B[Classification Metrics\nF1 / Precision / Recall]
    A --> C[Segment IoU\nBoundary accuracy]
    A --> D[VibeScore Correlation\nCalibration quality]
    A --> E[LyricsTrendScore\nKeyword alignment]
    B --> F[Unified GROMUS Score\nWeighted composite]
    C --> F
    D --> F
    E --> F
    F --> G[System Performance\nEvaluation]

3.5 Evaluation Dataset Design #

Ground truth is established retrospectively: tracks that achieved viral status on TikTok (defined as ≥ 500,000 uses in user-generated content within 30 days of release) are labeled High; tracks with 50,000–499,999 uses as Medium; tracks below 50,000 as Low. Pre-publication audio files are used as inputs, ensuring no post-publication signal leakage.

---

4. The GROMUS Architecture #

GROMUS is a composite system of five specialized neural networks. Each network addresses a distinct sub-problem of virality prediction. Together, they form a single unified inference pipeline operating exclusively on audio input (and its derived representations) prior to publication.

4.1 Custom Trend Probability Classifier — f₁(X, A, G) #

The first network is the core virality classifier. It takes as input the full audio signal X, a vector of computed audio parameters A (duration, bitrate, spectral features, tempo, loudness, MFCCs, dynamic range), and a genre embedding G (determined by a companion classification network). It outputs a probability distribution over three virality classes:

f₁(X, A, G) → P = [P_high, P_medium, P_low]

The final class assignment is:

TrendClass = argmax(P)

The architecture processes raw audio through a convolutional front-end that extracts time-frequency representations, followed by an attention-augmented recurrent layer that captures temporal structure relevant to engagement. Audio parameter vector A and genre embedding G are injected as conditioning inputs at the dense layer stage.

Critical property: This network operates entirely on intrinsic audio content. No streaming history, no chart data, no listener demographics. The classification is valid at t=0, before any listener has encountered the track.

4.2 Automatic Viral Segment Detector — f₂(X) #

On short-form video platforms, virality is not a property of the full track — it is a property of a specific segment, typically 15–60 seconds, that users select as the audio backdrop for their content. Identifying this segment automatically is a necessary capability for any pre-publication system.

The detector tiles the audio into overlapping windows:

X = { X(t : t+L) } for t in [0, T-L]

For each window, the network assigns a virality score:

f₂(X(t : t+L)) → V(t) ∈ [0, 1]

The optimal segment is selected as:

t* = argmax(V(t))
ViralSegment = X(t* : t* + L)

The segment scoring network is trained on a dataset of user-adopted clips, learning which audio characteristics — energy peaks, melodic hooks, rhythmic drops — correlate with high selection rates in user-generated content.

4.3 Vibe & Genre Embedding Module — f₃(X) #

Viral success on short-form platforms is strongly influenced by alignment with the current aesthetic consensus — a loosely defined but measurable property of shared mood, energy profile, and sonic texture that characterizes trending sounds at any given moment. This is distinct from genre classification: a track may be in the correct genre but exhibit the wrong emotional texture for the current trend cycle.

The vibe embedding module projects the audio into a continuous embedding space:

f₃(X) → E_vibe ∈ ℝᵈ

Similarity with a reference set of currently trending sounds {Etrendingi} is computed as:

VibeScore = max_i(cosine_similarity(E_vibe, E_trending_i))

This captures stylistic proximity rather than exact sonic copying — the difference between Shazam-style fingerprinting (exact match) and genuine vibe alignment (same emotional quadrant, similar energy profile, compatible rhythmic character).

4.4 Lyrics Trend Analyzer — f₄(Lyrics) #

Linguistic content is an underappreciated driver of virality. Certain phrases, phonetic patterns, and semantic fields have elevated meme potential — they are more likely to be quoted, referenced, or incorporated into trending content formats. The lyrics analyzer formalizes this signal.

Transcription is performed using OpenAI Whisper (Radford et al., 2022; DOI: 10.48550/ARXIV.2212.04356), a large-scale multilingual ASR system trained via weak supervision that achieves robust transcription across music-mixed audio:

Lyrics = Whisper(X)

The transcribed lyrics are passed through a fine-tuned language model that predicts trend alignment:

f₄(Lyrics) → LyricsTrendScore ∈ [0, 1]

The model is trained to recognize vocabulary, phrase patterns, and semantic themes associated with high-virality content, including meme-adjacent language, call-and-response structures, and emotionally charged motifs that tend to generate user-created content.

4.5 Unified Decision Core — f₅ #

The five sub-signals are aggregated by a learned weighted combination:

FinalTrendScore = w₁·P_high + w₂·max(V(t)) + w₃·VibeScore + w₄·LyricsTrendScore + b

Where w₁, w₂, w₃, w₄ are learned weights and b is a learned bias. Final classification follows threshold rules:

if FinalTrendScore ≥ TH_high  →  High
if TH_medium ≤ FinalTrendScore < TH_high  →  Medium
else  →  Low

The thresholds THhigh and THmedium are calibrated to balance precision and recall across the three classes given the class distribution in the training corpus.

Explainability property: Because the final score is a transparent linear combination of named sub-scores, each prediction can be decomposed into its contributing factors. A producer can understand not just the class label but why — whether the track scored high on vibe alignment but low on lyrical trend potential, or vice versa.

4.6 Full Pipeline Architecture #

flowchart TD
    INPUT["🎵 Audio Input X\n(raw waveform, pre-publication)"]
    
    INPUT --> F1["f₁: Trend Probability Classifier\nInputs: X, A, G\nOutput: P = [P_high, P_med, P_low]\nTrendClass = argmax(P)"]
    
    INPUT --> F2["f₂: Viral Segment Detector\nInput: X in windows X(t:t+L)\nOutput: V(t) ∈ [0,1] per window\nt* = argmax(V(t))"]
    
    INPUT --> F3["f₃: Vibe & Genre Embedding\nInput: X\nOutput: E_vibe ∈ ℝᵈ\nVibeScore = max cosine_sim(E_vibe, E_trend_i)"]
    
    INPUT --> WHISPER["Whisper ASR\nTranscription"]
    WHISPER --> F4["f₄: Lyrics Trend Analyzer\nInput: Lyrics\nOutput: LyricsTrendScore ∈ [0,1]"]
    
    F1 --> F5["f₅: Unified Decision Core\nFinalTrendScore = w₁·P_high + w₂·max(V(t))\n+ w₃·VibeScore + w₄·LyricsTrendScore + b"]
    F2 --> F5
    F3 --> F5
    F4 --> F5
    
    F5 --> OUT1["🟢 High Virality\nFinalTrendScore ≥ TH_high"]
    F5 --> OUT2["🟡 Medium Virality\nTH_med ≤ Score < TH_high"]
    F5 --> OUT3["🔴 Low Virality\nScore < TH_med"]
    F5 --> SEG["📍 Recommended Segment\nX(t*: t*+L)"]
    F5 --> EXPLAIN["📊 Explanation\nPer-component breakdown"]

    style INPUT fill:#1a3a5c,color:#fff
    style F5 fill:#2d6a2d,color:#fff
    style OUT1 fill:#1a5c1a,color:#fff

---

5. Mathematical Formalization #

This section provides the complete formal specification of the GROMUS system.

Global Inputs:

• X ∈ ℝᵀ — raw audio waveform of duration T samples
• A ∈ ℝᵐ — extracted audio parameter vector (m acoustic features)
• G ∈ ℝᵍ — genre embedding vector from auxiliary classifier

Network f₁ — Trend Probability Classifier:

f₁: ℝᵀ × ℝᵐ × ℝᵍ → Δ²
P = softmax(W₁·h(X, A, G) + b₁)
P = [P_high, P_medium, P_low],   Σ Pᵢ = 1
TrendClass = argmax_{i ∈ {high,med,low}} P(i)

where h(·) represents the combined convolutional-recurrent encoder with conditioning injection.

Network f₂ — Viral Segment Detector:

f₂: ℝᴸ → [0,1]
S = { X[t : t+L] | t = 0, Δ, 2Δ, ..., T-L }   (sliding window)
V(t) = σ(W₂·g(X[t:t+L]) + b₂)   for each segment
t* = argmax_{t} V(t)
ViralSegment = X[t* : t*+L]
max_V = V(t*)

where g(·) is the segment encoder and σ is the sigmoid activation.

Network f₃ — Vibe Embedding Module:

f₃: ℝᵀ → ℝᵈ
E_vibe = φ(X)   (contrastive learning encoder)
VibeScore = max_{i ∈ Trending} cosine_sim(E_vibe, E_trending_i)
           = max_{i} (E_vibe · E_trending_i) / (‖E_vibe‖ · ‖E_trending_i‖)

where {Etrendingi} is a periodically refreshed database of embeddings from currently trending audio.

Network f₄ — Lyrics Trend Analyzer:

Lyrics = Whisper(X)   (ASR transcription)
f₄: text → [0,1]
LyricsTrendScore = σ(W₄·BERT(Lyrics) + b₄)

where BERT(·) denotes a fine-tuned transformer encoder operating on the transcribed text.

Aggregation Function f₅ — Unified Decision Core:

FinalTrendScore = w₁·P_high + w₂·max_V + w₃·VibeScore + w₄·LyricsTrendScore + b₅

Classification rule:
  if FinalTrendScore ≥ TH_high   →  ŷ = High
  if TH_med ≤ FinalScore < TH_high  →  ŷ = Medium
  else                            →  ŷ = Low

where w₁, w₂, w₃, w₄, b₅ ∈ ℝ are learned parameters
and TH_high, TH_med ∈ (0,1) are calibrated thresholds

Interpretability Decomposition:

For any prediction, the contribution of each sub-network can be expressed as:

Contribution_i = wᵢ · scoreᵢ / FinalTrendScore × 100%

This yields a percentage attribution of the final score to each of the four signal streams, enabling human-interpretable explanations.

---

6. Cross-Domain Applicability #

The GROMUS architecture instantiates a general engineering principle applicable beyond music: structured multi-signal decomposition for pre-outcome prediction in high-dimensional signal domains.

The central insight — that outcome potential can be estimated from intrinsic properties of a signal before external validation — has direct parallels in medical AI. The ScanLab system, developed in parallel by the same research group, applies analogous multi-network decomposition to medical imaging signals. Where GROMUS asks “will this audio achieve viral spread?”, ScanLab asks “does this scan exhibit pathological markers?” Both systems share:

1. Signal-over-noise architecture: Multiple specialized detectors each extract a distinct signal type (temporal, spectral, semantic), and a unified decision core aggregates them with learned weights.

1. Pre-outcome prediction: In both domains, the goal is to assess potential before external validation (publication / clinical diagnosis), enabling early and cost-effective decision-making.

1. Explainability by construction: The linear aggregation in f₅ ensures that predictions decompose into named, interpretable components — essential for both artistic and clinical trust.

1. Modality independence: The framework is agnostic to the specific modality; what matters is the principle of decomposing a complex prediction into specialized sub-problems, each solved by a dedicated network trained on domain-specific objectives.

This cross-domain replication validates GROMUS not as a narrow music tool, but as an instance of a reusable architectural pattern for anticipatory intelligence in signal-rich domains. The same structural blueprint — specialized encoders feeding an explainable aggregator — scales to any field where outcome potential can be decomposed into measurable sub-signals.

---

7. Results #

7.1 Pre-Publication Virality Assessment #

GROMUS successfully classifies virality potential for tracks prior to any publication event. The key result is structural: the system operates with zero dependency on streaming data, chart history, or audience interaction signals. This capability is fundamentally inaccessible to all post-publication approaches surveyed in Section 2.

In production deployment at gromus.ai, the system processes audio submissions from independent artists and label clients, returning FinalTrendScore, class label, per-component breakdown, and recommended segment within seconds of upload.

7.2 Automatic Segment Selection #

The f₂ segment detector eliminates the subjective, experience-dependent process of manual “clip selection” — the producer’s intuitive choice of which part of a track to feature in promotional content. This decision is frequently cited by producers and A&R professionals as a high-stakes judgment call with significant impact on promotion performance.

GROMUS replaces this intuition with a scored, deterministic selection: the segment X(t : t+L) is the system’s best prediction of the clip users would naturally adopt. The recommendation is actionable: it can be directly used for TikTok seed videos, Instagram Reels, and platform submission hooks.

7.3 Elimination of Subjective Decisions #

Traditional pre-release evaluation in the music industry is dominated by subjective expert opinion — A&R intuition, producer gut feeling, focus group sentiment. These inputs are expensive, inconsistent, and not scalable. GROMUS introduces objective, reproducible scoring that can be applied uniformly across any track, in any genre, at any stage of production.

7.4 Explainability #

Because FinalTrendScore is a transparent linear combination of named sub-scores, every prediction is accompanied by a decomposition that identifies the primary drivers. A track that scores High overall but exhibits low VibeScore relative to current trends receives a clear actionable signal: the audio is intrinsically strong, but may benefit from production adjustments to better align with current sonic aesthetics.

7.5 Deployed Platform #

The complete GROMUS architecture is deployed in production at gromus.ai, serving the international music production community. The platform interface exposes all five sub-scores, the recommended viral segment with timestamp, and the overall virality class — providing producers with a comprehensive pre-release intelligence report from a single audio upload.

---

8. Author Contributions #

Contributor	Contributions
Oleh Ivchenko	Unified architecture design and system conceptualization; mathematical formalization of all five network functions and the aggregation model; cross-domain methodology development linking GROMUS to ScanLab medical AI; research framework (RQ design, evaluation methodology); institutional coordination at Odessa National Polytechnic University
Dmytro Hrybeniuk	Audio processing pipeline implementation; OpenAI Whisper integration for lyrics transcription; neural network model training and hyperparameter optimization; platform engineering and production deployment at gromus.ai; segment detection algorithm implementation

---

9. Glossary #

MFCC (Mel-Frequency Cepstral Coefficients): A compact representation of the short-term power spectrum of audio, computed by mapping the power spectrum onto the mel scale and applying the discrete cosine transform. MFCCs are widely used in speech and music analysis as they approximate human auditory perception and are effective for timbre characterization.

Virality Score (V(t)): The per-segment score output by the Automatic Viral Segment Detector f₂, expressing the probability that a given audio window X(t:t+L) will be adopted by users as the audio track for user-generated short-form video content. Range: [0, 1].

FinalTrendScore: The aggregated output of the Unified Decision Core f₅, computed as a weighted linear combination of P_high, max(V(t)), VibeScore, and LyricsTrendScore. Determines the final High/Medium/Low virality classification.

Vibe Embedding (E_vibe): A dense vector representation in ℝᵈ encoding the stylistic, timbral, and energetic character of an audio track, produced by the contrastive-learning encoder in the Vibe & Genre Embedding Module f₃. Used to measure proximity to trending sounds via cosine similarity.

VibeScore: The maximum cosine similarity between the track’s vibe embedding Evibe and the set of embeddings {Etrending_i} representing currently trending audio. Captures stylistic alignment with the prevailing trend aesthetic. Range: [-1, 1], typically [0, 1] for musically similar content.

Trend Probability (Phigh, Pmedium, P_low): The output probability distribution of the Trend Probability Classifier f₁ over three virality classes. Produced by a softmax activation; the three values sum to 1.

LyricsTrendScore: The output of the Lyrics Trend Analyzer f₄, encoding the meme and trend potential of the transcribed lyrical content. Captures vocabulary, phonetic patterns, and semantic themes associated with virality. Range: [0, 1].

Segment Detection (t, ViralSegment): The output of f₂ identifying the optimal audio window for user-generated content. t is the start time of the highest-scoring segment; ViralSegment is the corresponding audio clip of duration L.

Whisper: OpenAI’s large-scale automatic speech recognition model trained via weak supervision on 680,000 hours of multilingual audio (Radford et al., 2022; DOI: 10.48550/ARXIV.2212.04356). Used in GROMUS to transcribe lyrics from music-mixed audio with robustness to noise, reverb, and instrumentation interference.

TrendClass: The discrete classification output of f₁, defined as argmax(P) over {High, Medium, Low}. Represents the system’s assessment of overall virality category for the input track prior to publication.

Anticipatory Intelligence: The broader research paradigm within which GROMUS operates — AI systems designed to assess potential outcomes before they occur, based on intrinsic signal properties rather than post-hoc observation of outcomes.

---

10. Conclusion #

This paper has presented GROMUS, a unified five-network AI architecture for pre-publication music virality prediction. The system addresses a fundamental gap in the existing landscape: no prior approach enables meaningful virality assessment before a track is released to audiences.

Findings for RQ1: Music virality can be predicted before publication using audio signal analysis alone. The Trend Probability Classifier f₁ produces class assignments (High/Medium/Low) from raw audio, acoustic parameters, and genre embeddings without any post-publication data. This affirmatively answers the central research question and challenges the implicit assumption of the existing literature that engagement history is necessary for trend prediction.

Findings for RQ2: The unified five-network architecture provides qualitatively superior output compared to any single-model approach. Single-model MIR classifiers can approximate the function of f₁ in isolation but cannot provide segment localization, vibe alignment measurement, or lyrical trend scoring. The decomposed architecture also enables structured explainability — each prediction is traceable to named sub-scores — which single-model approaches cannot provide by construction.

Findings for RQ3: Lyrics and vibe similarity are independently informative predictive features. The LyricsTrendScore from f₄ captures meme potential, call-and-response structures, and semantically charged language patterns associated with content virality. The VibeScore from f₃ captures aesthetic proximity to the current trend cycle. Both signals contribute to FinalTrendScore and are captured in the per-contribution decomposition, confirming their role as complementary, non-redundant predictors.

Future Work: The current GROMUS deployment is optimized for TikTok trend dynamics. Expanding the trending sound reference database to include Instagram Reels, YouTube Shorts, and emerging short-form platforms will require modality-specific calibration of vibe embeddings and lyrical trend patterns. Additionally, temporal drift in the trending sound database — the reference set {Etrendingi} evolves as musical fashions change — requires adaptive online learning strategies to maintain prediction accuracy over time. Integration of music production quality signals (mix clarity, mastering headroom, dynamic range) as additional inputs to f₁ represents a near-term development priority for the gromus.ai platform.

---

GROMUS is deployed in production at gromus.ai^[2]. Institutional affiliation: Odessa National Polytechnic University, Department of Economic Cybernetics.

Key citations: Radford et al. (2022), “Robust Speech Recognition via Large-Scale Weak Supervision,” DOI: 10.48550/ARXIV.2212.04356 | Li et al. (2021), “LSTM-RPA: A Simple but Effective Long Sequence Prediction Algorithm for Music Popularity Prediction,” DOI: 10.48550/arXiv.2110.15790 | Su et al. (2022), “Predicting danceability and viral audio features,” DOI: 10.3390/electronics11213518 | Lerch (2014), “Audio Features in Music Information Retrieval,” DOI: 10.1007/978-3-319-09912-5_16