GoogleAudio modelPreview
Lyria Realtime
PopularityRank 591 of 876
19out of 100same as last month
Summary
Lyria Realtime is Google’s experimental model for making music while the creative controls change. It accepts weighted text prompts and live control parameters, then streams music audio in real time. The catalog emphasizes high fidelity and granular creative control. Lyria Realtime streams changing music as control input arrives, rather than producing a finished background track from one prompt.
Lyria 3.5 generates full songs and Lyria 3 Clip short clips. Lyria Realtime instead changes the music as control input arrives, keeping it functionally separate from both batch variants.
All benchmark results
Every published result with its source, test setup and comparison models.
Show all 9 published benchmark results
Measurements without matching peer values
These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.
MazzikaAI system-level Bayati grounding: voiced-frame off-grid share (Show measurement, test conditions, and source)
- Source value
- 22.4
- Score
- 22.4
- Metric
- Share of pYIN predominant-F0 voiced frames at least 35 cents from the nearest 12-tone equal-tempered semitone; targeted Arabic maqam bayati output, not a general quality metric
- Unit
- %
- Benchmark version
- arXiv:2608.10360v1 (2026-08-11), Section 6.4, Figure 6
- Category
- audio
- Direction
- Higher is better
- Lower confidence bound
- 21.7
- Upper confidence bound
- 23.2
- Source type
- independent-evaluator
- Evaluator
- MazzikaAI, arXiv:2608.10360v1
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. System-level MazzikaAI run using the exact Google Gemini API endpoint models/lyria-realtime-exp; Arabic maqam bayati. Grounding condition A uses the explicit maqam-grounding clause. Three input-identical replay arms driven by the same recorded primary-session performance, about 18 minutes of generated audio and 553 melody chunks. This measures system-plus-prompt-controller output, not the isolated base model.
- Context
- This is a system-level result, not an unprompted or general Lyria quality score: MazzikaAI steers the exact models/lyria-realtime-exp endpoint with explicit Arabic maqam bayati prompts. More off-grid voiced frames is directional only for this microtonal target; it is not a general quality score. The paper uses one author's recorded performance as replay input and reports qualitative playing-session feedback from two expert musicians, not a controlled larger listening benchmark. The frame denominator and 95% CI method are not stated. No Lyria 3 or Lyria 3 Pro scores are used.
MazzikaAI system-level Bayati grounding: voiced-frame off-grid share (Show measurement, test conditions, and source)
- Source value
- 14.9
- Score
- 14.9
- Metric
- Share of pYIN predominant-F0 voiced frames at least 35 cents from the nearest 12-tone equal-tempered semitone; matched-prompt ablation for Arabic maqam bayati
- Unit
- %
- Benchmark version
- arXiv:2608.10360v1 (2026-08-11), Section 6.4, Figure 6
- Category
- audio
- Direction
- Higher is better
- Lower confidence bound
- 14
- Upper confidence bound
- 15.7
- Source type
- independent-evaluator
- Evaluator
- MazzikaAI, arXiv:2608.10360v1
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. System-level MazzikaAI run using the exact Google Gemini API endpoint models/lyria-realtime-exp; Arabic maqam bayati. Matched ablation condition B replaces the maqam-grounding clause with a same-length generic Western-scale phrase. Three input-identical replay arms driven by the same recorded primary-session performance, about 18 minutes of generated audio and 521 melody chunks. This measures system-plus-prompt-controller output, not the isolated base model.
- Context
- This is a system-level result, not an unprompted or general Lyria quality score: MazzikaAI steers the exact models/lyria-realtime-exp endpoint; condition B replaces the Bayati grounding clause with a matched-length generic Western-scale phrase. More off-grid voiced frames is directional only for this microtonal target; it is not a general quality score. The paper uses one author's recorded performance as replay input and reports qualitative feedback from two expert musicians, not a controlled larger listening benchmark. The frame denominator and 95% CI method are not stated. No Lyria 3 or Lyria 3 Pro scores are used.
MazzikaAI system-level Bayati grounding: half-flat second-degree frame share (Show measurement, test conditions, and source)
- Source value
- 59.4
- Score
- 59.4
- Metric
- Share of pYIN predominant-F0 frames in the 125-175 cent E-koron half-flat band when melody visits the second-degree region above D
- Unit
- %
- Benchmark version
- arXiv:2608.10360v1 (2026-08-11), Section 6.4, Figure 6
- Category
- audio
- Direction
- Higher is better
- Sample
- 650
- Source type
- independent-evaluator
- Evaluator
- MazzikaAI, arXiv:2608.10360v1
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. System-level MazzikaAI run using the exact Google Gemini API endpoint models/lyria-realtime-exp; Arabic maqam bayati. Grounding condition A uses the explicit maqam-grounding clause. Three input-identical replay arms driven by the same recorded primary-session performance, about 18 minutes of generated audio and 553 melody chunks. The reported N is the 650 pYIN frames in the second-degree-region denominator for this condition. This measures system-plus-prompt-controller output, not the isolated base model.
- Context
- This is a system-level, prompt-conditioned microtonal output measure, not a general music-quality result or isolated-model score. The source says that aggregate second-degree CQT energy did not show a concentrated 150-cent peak and that energy split nearly evenly across E-flat, E-half-flat and E in both conditions; prompt grounding increased the melodic line's measured occupancy of the half-flat band but did not provide stable tuning control. One author's recorded performance and three stochastic replay arms limit generalization. No Lyria 3 or Lyria 3 Pro scores are used.
MazzikaAI system-level Bayati grounding: half-flat second-degree frame share (Show measurement, test conditions, and source)
- Source value
- 24.1
- Score
- 24.1
- Metric
- Share of pYIN predominant-F0 frames in the 125-175 cent E-koron half-flat band when melody visits the second-degree region above D; matched-prompt ablation
- Unit
- %
- Benchmark version
- arXiv:2608.10360v1 (2026-08-11), Section 6.4, Figure 6
- Category
- audio
- Direction
- Higher is better
- Sample
- 1530
- Source type
- independent-evaluator
- Evaluator
- MazzikaAI, arXiv:2608.10360v1
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. System-level MazzikaAI run using the exact Google Gemini API endpoint models/lyria-realtime-exp; Arabic maqam bayati. Matched ablation condition B replaces the maqam-grounding clause with a same-length generic Western-scale phrase. Three input-identical replay arms driven by the same recorded primary-session performance, about 18 minutes of generated audio and 521 melody chunks. The reported N is the 1,530 pYIN frames in the second-degree-region denominator for this condition. This measures system-plus-prompt-controller output, not the isolated base model.
- Context
- This is a system-level, prompt-conditioned microtonal output measure, not a general music-quality result or isolated-model score. Condition B replaces the Bayati grounding clause with a matched-length generic Western-scale phrase. The source says aggregate second-degree CQT energy did not show a concentrated 150-cent peak and split nearly evenly across E-flat, E-half-flat and E in both conditions. One author's recorded performance and three stochastic replay arms limit generalization. No Lyria 3 or Lyria 3 Pro scores are used.
MazzikaAI system-level Bayati grounding: tonic-D CQT energy share (Show measurement, test conditions, and source)
- Source value
- 0.14
- Score
- 0.14
- Metric
- Pooled constant-Q energy share assigned to the tonic D pitch class across three replay arms
- Unit
- ratio
- Benchmark version
- arXiv:2608.10360v1 (2026-08-11), Section 6.4, Figure 6
- Category
- audio
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- MazzikaAI, arXiv:2608.10360v1
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. System-level MazzikaAI run using the exact Google Gemini API endpoint models/lyria-realtime-exp; Arabic maqam bayati. Grounding condition A uses the explicit maqam-grounding clause. Constant-Q energy profile at 10-cent resolution, pooled across three input-identical replay arms driven by the same recorded primary-session performance. This measures system-plus-prompt-controller output, not the isolated base model.
- Context
- This is a narrow system-level tonal-grounding measure, not a general music-quality score. The pooled value hides substantial stochastic variation: one of the three grounded arms failed to anchor D, ranking it sixth. A larger D share is only directionally favorable for this particular tonic-D grounding objective. One author's recorded performance is the replay input; this is not a controlled larger listening benchmark. No Lyria 3 or Lyria 3 Pro scores are used.
MazzikaAI system-level Bayati grounding: tonic-D CQT energy share (Show measurement, test conditions, and source)
- Source value
- 0.094
- Score
- 0.094
- Metric
- Pooled constant-Q energy share assigned to the tonic D pitch class across three replay arms
- Unit
- ratio
- Benchmark version
- arXiv:2608.10360v1 (2026-08-11), Section 6.4, Figure 6
- Category
- audio
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- MazzikaAI, arXiv:2608.10360v1
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. System-level MazzikaAI run using the exact Google Gemini API endpoint models/lyria-realtime-exp; Arabic maqam bayati. Matched ablation condition B replaces the maqam-grounding clause with a same-length generic Western-scale phrase. Constant-Q energy profile at 10-cent resolution, pooled across three input-identical replay arms driven by the same recorded primary-session performance. This measures system-plus-prompt-controller output, not the isolated base model.
- Context
- This is a narrow system-level tonal-grounding measure, not a general music-quality score. Condition B replaces the Bayati-grounding clause with a same-length generic Western-scale phrase. The pooled D share is 0.094 and ranks fourth; D never ranks first in any of the three ablation arms. One author's recorded performance is the replay input; this is not a controlled larger listening benchmark. No Lyria 3 or Lyria 3 Pro scores are used.
MazzikaAI note-triggered keypress-to-audible latency, median (Show measurement, test conditions, and source)
- Source value
- 263
- Score
- 263
- Metric
- Direct end-to-end system latency from causal key press to first new-prompt chunk scheduled as audible, median
- Unit
- ms
- Benchmark version
- arXiv:2608.10360v1, 2026-08-11, section 6.1, Table 5
- Direction
- Lower is better
- Sample
- 98
- Source type
- independent-evaluator
- Evaluator
- MazzikaAI evaluation authors
- Status
- active
- Retrieved at
- 2026-10-24
- Methodology
- Independent evaluation. MazzikaAI live browser/FastAPI system using the exact Google Gemini API endpoint models/lyria-realtime-exp. Primary 8.5-minute live blues session, 1,516 note events, both engines active. Directly measured note-triggered re-prompts: median 263 ms, p95 818 ms, p99 945 ms, n=98 latency events. This is system-event N, not model test-case N.
- Context
- System-level latency, not isolated model inference time. It includes controller, API, transport, and client scheduling. The paper describes one primary session and reports no model checkpoint revision or confidence interval.
MazzikaAI melody prompt-push-to-next-chunk latency, median (Show measurement, test conditions, and source)
- Source value
- 150
- Score
- 150
- Metric
- Latency from melody prompt push to arrival of the next chunk, median
- Unit
- ms
- Benchmark version
- arXiv:2608.10360v1, 2026-08-11, section 6.1, Table 5
- Direction
- Lower is better
- Sample
- 140
- Source type
- independent-evaluator
- Evaluator
- MazzikaAI evaluation authors
- Status
- active
- Retrieved at
- 2026-10-24
- Methodology
- Independent evaluation. MazzikaAI live browser/FastAPI system using models/lyria-realtime-exp. Primary 8.5-minute live blues session. The melody push-to-next-chunk row reports median 150 ms, p95 1,070 ms, p99 1,799 ms, n=140 latency events. This is system-event N, not model test-case N.
- Context
- System-level endpoint measurement, not isolated model inference. The first chunk after prompt push is only a lower bound for when prompt-affected audio arrives because chunk content was not analyzed for prompt attribution. No model checkpoint revision or confidence interval is reported.
MazzikaAI steady-state melody chunk inter-arrival, median (Show measurement, test conditions, and source)
- Source value
- 1999
- Score
- 1999
- Metric
- Server-side inter-arrival time between consecutive melody chunk emission records, median
- Unit
- ms
- Benchmark version
- arXiv:2608.10360v1, 2026-08-11, section 6.1, Figure 4 and Table 5
- Direction
- No direction published
- Sample
- 142
- Source type
- independent-evaluator
- Evaluator
- MazzikaAI evaluation authors
- Status
- active
- Retrieved at
- 2026-10-24
- Methodology
- Independent evaluation. MazzikaAI system using models/lyria-realtime-exp; melody chunks are 2.0 seconds long. In the primary 8.5-minute session, median cadence is 1,999 ms, p95 2,978 ms, p99 3,459 ms, n=142 of 146 gaps after excluding four gaps of at least five seconds spanning pauses or restarts. This is system-event N, not model test-case N.
- Context
- System-level server emission cadence, not a general model-speed guarantee. Four of 146 melody gaps of at least five seconds were excluded because they spanned pauses or stream restarts. No confidence interval or checkpoint revision is reported.
Popularity compared
How often the model is searched for, used and mentioned, compared with every model of its category.
Popularity index
0 to 100, higher is more popular
20 of 83 models
As of November 2026 · Source: gradually.ai popularity index
Over time
Weekly values from October 18, 2026 to November 6, 2026
Lyria RealtimeIndex-Echo-S2TT-2B 26Index-Echo-S2TT-9B 24Eleven Flash v2.5 22Voxtral Mini Transcribe Realtime 18
Method changes in the timeline, a jump at these points is no market event:
- 1November 1, 2026: Value only with search or usage data, 30-day attention
What is gaining popularity right now, across models, vendors, harnesses and AI subscriptions, is collected in All AI trends.
Similar models
More models of the same kind as Lyria Realtime, chosen by model line, developer, size, price and release date.
- Gemini 3.5 TranscribeGoogle · same developer
- Lyria 3.5Google · same developer
- Gemini 3.5 Live TranslateGoogle · same developer
- Gemini 3.1 Flash TTSGoogle · same developer
- Gemini 3.5 Transcribe LiveGoogle · same developer
- Lyria 3 ClipGoogle · same developer
- Gemini 3.8 LiveGoogle · same developer
- Gemini 3.8 Live Extended ThinkingGoogle · same developer
All specifications
Architecture
Input and output
Watermarking and provenance
AI watermarking refers to invisible signals embedded in the output. Visible marks and provenance metadata are listed separately. Unknown means there is insufficient evidence for that particular field.
- AI watermarking
- Yes. Google documents SynthID for audio generated or published with Lyria and lists Lyria RealTime as a model in that family.
- Source 1: AI watermarkingSource 2: AI watermarking
Licence and access
API
Further details
Sources and data date
Show all 13 sources
Benchmark sources at a glance
Which independent evaluators tested this model. This is coverage, not a quality ranking.
No active result from the reviewed external evaluators can currently be attributed to this model. Other source evidence remains available below.
Without vendor claims and historical or disputed results; those stay labelled in the single values.
| Type | Evidence and data date |
|---|---|
| Research status | Research date October 15, 2026. 6 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication. |
| Unresolved | quality-benchmarks Note The MazzikaAI paper reports six targeted Bayati measurements from a prompt-and-controller system steering the exact lyria-realtime-exp route. The rows cover voiced-frame pitch grounding, half-flat frame share, and pooled tonic-D energy, including matched-prompt comparisons. These are system-plus-controller replay measurements from one recorded performance, not isolated Lyria scores or general music-quality results. Published frame counts and confidence intervals remain limited to the individual rows, and the paper does not establish broad audio quality. |
| Not found in the inspected sources | quality-benchmarks Note No exact numeric quality score for lyria-realtime-exp was found in the reviewed official documentation or current independent music leaderboards. The DeepMind paper's numeric Table 1 results are explicitly for Magenta RealTime, a different model, so none are transferred. |
| Not found in the inspected sources | quality-benchmarks Note No exact independent quality score was found for Lyria Realtime experimental. The official page states a maximum control latency of two seconds, but that is not a musical-quality score; completed-track Lyria results do not transfer to the real-time model. |
| Not published | quality-benchmarks Note The exact Google API model documentation publishes no measured quality score, evaluation cohort, or result denominator. Model capability and technical audio specifications are not benchmark results. |
| Additional source | Google Gemini API model catalog (retrieved September 22, 2026) · Google Gemini API model catalog · Editorial description reviewed October 2, 2026 |
| Additional source | MazzikaAI, arXiv:2608.10360v1 (retrieved October 17, 2026; October 24, 2026) |
| Additional source | Model metadata (retrieved October 15, 2026; October 16, 2026) |
| Additional source | Model metadata (retrieved October 15, 2026) |
| Additional source | Model metadata (retrieved October 16, 2026) |
| Additional source | Google DeepMind SynthID (retrieved November 5, 2026) |
| Additional source | Google DeepMind Lyria (retrieved November 5, 2026) |
| Additional source | OpenRouter: Popularity index (derived), CC BY 4.0 |
| Additional source | Vercel AI Gateway: Popularity index (derived), CC BY 4.0 |
| Additional source | Ollama: cumulative downloads for the popularity index |
| Additional source | Replicate: cumulative runs for the popularity index |
| Additional source | Google Trending Now: limited feed selection for Germany and the US |
| Additional source | Creative Commons Attribution 4.0 |