hexgradAudio modelActive
Kokoro-82M
PopularityRank 98 of 876
51out of 100same as last month
Summary
Kokoro-82M is a local open-weight TTS checkpoint with 82 million parameters. The weights use Apache 2.0, and v1.0 lists eight languages, 54 voices, and a 24 kHz sample rate. The architecture uses StyleTTS 2 and ISTFTNet in a decoder-only setup without a released diffusion or encoder component.
All benchmark results
Every published result with its source, test setup and comparison models.
Show all 81 published benchmark results
Page 1 of 7
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.
Paradee paper Kokoro-82M teacher Whisper WER (af_heart) (Show measurement, test conditions, and source)
- Source value
- 5.7
- Score
- 5.7
- Metric
- Whisper base word error rate against source text
- Unit
- %
- Benchmark version
- Paradee paper arXiv:2610.06817v1, Table 2, 200 held-out sentences
- Category
- audio
- Direction
- Lower is better
- Sample
- 200
- Source type
- independent-evaluator
- Evaluator
- Paradee authors, arXiv
- Status
- active
- Retrieved at
- 2026-10-16
- Methodology
- Independent evaluation. Exact Kokoro-82M v1.0 teacher, voice af_heart, evaluated on 200 held-out English sentences from the paper's WikiText-103 synthesis corpus. The paper transcribes generated audio with Whisper base and compares the transcript with source text. The paper table does not give a WER confidence interval value.
- Context
- This is an automated Whisper-based intelligibility check, not a human transcript-rating panel. It is specific to one voice and this English held-out set. The paper reports a WER point value but no separate WER confidence interval, so none is recorded.
Artificial Analysis Provider Voice Arena preference (Show measurement, test conditions, and source)
- Source value
- 1061
- Score
- 1061
- Metric
- blind pairwise native-voice preference Elo
- Unit
- elo
- Benchmark version
- Current Provider Voice Arena board retrieved 2026-10-15
- Category
- audio
- Direction
- Higher is better
- Sample
- 5361
- Confidence interval
- 11
- Lower confidence bound
- 1050
- Upper confidence bound
- 1072
- Source type
- independent-evaluator
- Evaluator
- Artificial Analysis
- Status
- superseded
- Retrieved at
- 2026-10-15
- Methodology
- Independent evaluation. Blind pairwise preference on provider-native voice set (8 voices); 5,361 board appearances; model release shown as Jan 2025.
- Context
- Appearances are not unique listeners. The current page does not state a per-model evaluation date range.
Paradee paper Kokoro-82M teacher UTMOS (af_heart) (Show measurement, test conditions, and source)
- Source value
- 4.52
- Score
- 4.52
- Metric
- UTMOS neural prediction of MOS for naturalness, using the utmos22_strong checkpoint
- Unit
- rating
- Benchmark version
- Paradee paper arXiv:2610.06817v1, Table 2, 200 held-out sentences
- Category
- audio
- Direction
- Higher is better
- Sample
- 200
- Confidence interval
- 0.003
- Source type
- independent-evaluator
- Evaluator
- Paradee authors, arXiv
- Status
- active
- Retrieved at
- 2026-10-16
- Methodology
- Independent evaluation. Exact Kokoro-82M v1.0 teacher, voice af_heart, evaluated on 200 held-out English sentences from the paper's WikiText-103 synthesis corpus. The paper uses the utmos22_strong neural MOS-prediction checkpoint; it reports a 95% confidence interval of plus or minus 0.003 for this teacher result.
- Context
- UTMOS is a neural predictor trained on human ratings, not a human listening score from this study. The result is specific to one voice and this 200-sentence English test set.
Paradee Table 4 Kokoro-82M teacher DTW log-mel floor (200 held-out sentences) (Show measurement, test conditions, and source)
- Source value
- 0.12
- Score
- 0.12
- Metric
- Dynamic-time-warped log-mel L1 distance to the teacher's second stochastic rendition
- Unit
- ratio
- Benchmark version
- Paradee arXiv:2610.06817v1, Table 4
- Category
- audio
- Direction
- Lower is better
- Sample
- 200
- Source type
- independent-evaluator
- Evaluator
- Paradee paper authors, arXiv
- Status
- active
- Retrieved at
- 2026-10-16
- Methodology
- Independent evaluation. Exact Kokoro-82M v1.0 teacher text side rendered through its frozen teacher decoder on 200 held-out English sentences. Table 4 defines this teacher-row DTW value as the teacher rendition compared with a second run using a different random seed; it is a stochastic-reference floor, not a cross-model quality rank.
- Context
- This is a paper-reported spectral fidelity diagnostic, not human preference or naturalness. It is specific to one teacher voice/configuration and the held-out English sentences; no confidence interval is reported.
Paradee Appendix Table 6 Kokoro-82M teacher self-run stochastic reference (Show measurement, test conditions, and source)
- Source value
- 0.09
- Score
- 0.09
- Metric
- Dynamic-time-warped log-mel L1 distance
- Unit
- ratio
- Benchmark version
- Paradee arXiv:2610.06817v1, Appendix Table 6
- Category
- audio
- Direction
- Lower is better
- Sample
- 5
- Source type
- independent-evaluator
- Evaluator
- Paradee paper authors, arXiv
- Status
- active
- Retrieved at
- 2026-10-16
- Methodology
- Independent evaluation. Exact Kokoro-82M teacher against its own second run with a different random excitation, on the same five English sentences used for the Appendix A DTW log-mel reference points. Lower means closer spectral alignment; this is a stochastic baseline, not an independent quality score.
- Context
- This five-sentence diagnostic estimates run-to-run variation and does not rank models. No confidence interval is reported.
Paradee Appendix Table 6 Kokoro-82M teacher full-weight fp16 reference (Show measurement, test conditions, and source)
- Source value
- 0.11
- Score
- 0.11
- Metric
- Dynamic-time-warped log-mel L1 distance from full-precision teacher
- Unit
- ratio
- Benchmark version
- Paradee arXiv:2610.06817v1, Appendix Table 6
- Category
- audio
- Direction
- Lower is better
- Sample
- 5
- Source type
- independent-evaluator
- Evaluator
- Paradee paper authors, arXiv
- Status
- active
- Retrieved at
- 2026-10-16
- Methodology
- Independent evaluation. Exact Kokoro-82M teacher, all weights stored in fp16, compared with the full-precision teacher on the paper's same five English sentences. Table 6 reports DTW log-mel distance; this is a precision/reference diagnostic, not a model-quality rank.
- Context
- This five-sentence result measures spectral fidelity to the full-precision teacher, not human preference or naturalness. No confidence interval is reported.
Paradee Appendix Table 6 Kokoro-82M teacher ONNX fp32 export fidelity (Show measurement, test conditions, and source)
- Source value
- 0.21
- Score
- 0.21
- Metric
- Dynamic-time-warped log-mel L1 distance from PyTorch teacher
- Unit
- ratio
- Benchmark version
- Paradee arXiv:2610.06817v1, Appendix Table 6
- Category
- audio
- Direction
- Lower is better
- Sample
- 5
- Source type
- independent-evaluator
- Evaluator
- Paradee paper authors, arXiv
- Status
- active
- Retrieved at
- 2026-10-16
- Methodology
- Independent evaluation. Exact Kokoro-82M teacher ONNX fp32 export compared against its PyTorch reference on the same five English sentences. This measures export/runtime output fidelity, not an independent model-quality rank.
- Context
- This five-sentence result measures export/runtime fidelity to the PyTorch implementation, not human preference or naturalness. No confidence interval is reported.
Paradee Appendix Table 6 Kokoro-82M teacher all-weight int8 per-channel fidelity (Show measurement, test conditions, and source)
- Source value
- 0.28
- Score
- 0.28
- Metric
- Dynamic-time-warped log-mel L1 distance from full-precision teacher
- Unit
- ratio
- Benchmark version
- Paradee arXiv:2610.06817v1, Appendix Table 6
- Category
- audio
- Direction
- Lower is better
- Sample
- 5
- Source type
- independent-evaluator
- Evaluator
- Paradee paper authors, arXiv
- Status
- disputed
- Retrieved at
- 2026-10-16
- Methodology
- Independent evaluation. Exact Kokoro-82M teacher with all weights int8 per-channel and activations kept fp32, compared with full precision on the same five English sentences. Appendix Table 6 reports 25 ms duration change. Table 7 reports a near-matching all-weight 8-bit value of 0.270 under its separately described per-part quantization experiment; both source values are retained as disputed, not reconciled.
- Context
- This five-sentence metric measures spectral fidelity to the full-precision teacher, not human preference or naturalness. The paper reports differing aggregate values across Tables 6 and 7, so this value is marked disputed; no confidence interval is given.
Paradee Appendix Table 6 Kokoro-82M teacher CPU versus Apple MPS output fidelity (Show measurement, test conditions, and source)
- Source value
- 0.39
- Score
- 0.39
- Metric
- Dynamic-time-warped log-mel L1 distance between CPU and Apple MPS runs
- Unit
- ratio
- Benchmark version
- Paradee arXiv:2610.06817v1, Appendix Table 6
- Category
- audio
- Direction
- Lower is better
- Sample
- 5
- Source type
- independent-evaluator
- Evaluator
- Paradee paper authors, arXiv
- Status
- active
- Retrieved at
- 2026-10-16
- Methodology
- Independent evaluation. Exact Kokoro-82M teacher, CPU run compared with Apple MPS output on the same five English sentences. This is a backend output-fidelity comparison, not a model-quality score.
- Context
- This five-sentence result measures implementation/backend differences, not human preference or naturalness. No confidence interval is reported.
Paradee Appendix Table 6 Kokoro-82M teacher prebuilt dynamic-int8 ONNX fidelity (Show measurement, test conditions, and source)
- Source value
- 0.58
- Score
- 0.58
- Metric
- Dynamic-time-warped log-mel L1 distance from full-precision teacher
- Unit
- ratio
- Benchmark version
- Paradee arXiv:2610.06817v1, Appendix Table 6
- Category
- audio
- Direction
- Lower is better
- Sample
- 5
- Source type
- independent-evaluator
- Evaluator
- Paradee paper authors, arXiv
- Status
- active
- Retrieved at
- 2026-10-16
- Methodology
- Independent evaluation. Exact Kokoro-82M teacher using the paper's prebuilt dynamic-int8 ONNX export, with activations quantized, compared with full-precision teacher output on the same five English sentences. Table 6 reports 65 ms duration change.
- Context
- This five-sentence result measures spectral fidelity to the full-precision teacher, not human preference or naturalness. No confidence interval is reported.
Paradee Appendix Table 6 Kokoro-82M teacher all-weight int4 group-64 fidelity (Show measurement, test conditions, and source)
- Source value
- 0.67
- Score
- 0.67
- Metric
- Dynamic-time-warped log-mel L1 distance from full-precision teacher
- Unit
- ratio
- Benchmark version
- Paradee arXiv:2610.06817v1, Appendix Table 6
- Category
- audio
- Direction
- Lower is better
- Sample
- 5
- Source type
- independent-evaluator
- Evaluator
- Paradee paper authors, arXiv
- Status
- disputed
- Retrieved at
- 2026-10-16
- Methodology
- Independent evaluation. Exact Kokoro-82M teacher with all weights int4 group-64, compared with full precision on the same five English sentences. Appendix Table 6 reports 60 ms duration change. Table 7 separately reports 0.624 for all weights in its 4-bit groups-of-64 column; both source values remain disputed and unreconciled.
- Context
- This five-sentence metric measures spectral fidelity to the full-precision teacher, not human preference or naturalness. The paper reports differing aggregate values across Tables 6 and 7, so this value is marked disputed; no confidence interval is given.
Paradee Appendix Table 6 Kokoro-82M af_bella voice-reference distance (Show measurement, test conditions, and source)
- Source value
- 1.27
- Score
- 1.27
- Metric
- Dynamic-time-warped log-mel L1 distance from af_heart teacher reference
- Unit
- ratio
- Benchmark version
- Paradee arXiv:2610.06817v1, Appendix Table 6
- Category
- audio
- Direction
- Lower is better
- Sample
- 5
- Source type
- independent-evaluator
- Evaluator
- Paradee paper authors, arXiv
- Status
- active
- Retrieved at
- 2026-10-16
- Methodology
- Independent evaluation. Exact Kokoro-82M v1.0 teacher rendering the same English text with voice af_bella, compared with the af_heart teacher reference on five English sentences. This deliberately changes the speaker voice, so it is a reference-sensitivity diagnostic, not a score for the af_heart configuration.
- Context
- This five-sentence diagnostic intentionally compares two different voices and must not be interpreted as same-voice model quality. No confidence interval is reported.
Page 1 of 7
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
Kokoro-82MConfucius4-R2T2 52NVIDIA Nemotron 3.5 Transcribe 52Granite Speech 5.0 470M TurboCTC 52Nova-3 51
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.
Kokoro-82M compared
Similar models
More models of the same kind as Kokoro-82M, chosen by model line, developer, size, price and release date.
- Granite Speech 5.0 470M TurboCTCIBM · similar size
- Granite Speech 5.0 470M TurboCTC NCIBM · similar size
- Fish Audio S2 ProFish Audio · released around the same time
- OrukeetOruk · similar size
- MOSS-Audio-4B-ThinkingOpenMOSS Team · released around the same time
- Nemotron-3-DiarizationNVIDIA · released around the same time
- pyannote Community-1pyannote · popular in this category
- NVIDIA Nemotron 3.5 TranscribeNVIDIA · released around the same time
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
- Unknown
- no source
- Watermarking technique
- Not known
- no source
Licence and access
API
Further details
Sources and data date
Show all 19 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. 14 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication. |
| Not published | per-model-denominator-and-confidence-interval Note The current Hume TTS page does not publish a per-model sample count or confidence interval for the frontier score. |
| Not published | benchmark-model-weight-revision Note The report names model repositories and runtime versions but does not pin the weight revision used. The current Hugging Face API revision cannot be represented as the historical run revision. |
| Unresolved | cpu-run-reconciliation Note The pinned July 6 repository report and the June 22 public article report different values for some cells; the reason and relationship between the runs are not published. |
| Unresolved | per-type N and interval definition Note The Dograh page omits per-type N except Date (40). Its Date detail labels 0.03–0.20 a “95% range” without defining the interval; no CI fields are populated. |
| Additional source | Kokoro model card (retrieved September 24, 2026; October 15, 2026; October 16, 2026; October 17, 2026) · Kokoro-82M model card · Editorial description reviewed October 2, 2026 |
| Additional source | Artificial Analysis (retrieved October 15, 2026; November 3, 2026) |
| Additional source | Paradee authors, arXiv (retrieved October 16, 2026) |
| Additional source | Hume AI Real World VoiceEQ (retrieved October 17, 2026) |
| Additional source | HeyNEO Kokoro CPU Benchmark (retrieved October 17, 2026) |
| Additional source | DLYog Lab Research Services LLC, Distribution-Time Voice Specialization technical report (retrieved October 18, 2026) |
| Additional source | Pakorn Nathong and Kunat Pipatanakul, Paxa Labs, Kokoro-82M CPU serving study (retrieved October 18, 2026) |
| Additional source | Dograh TTS verbalization benchmark (retrieved October 19, 2026) |
| Additional source | LiteRT Community Kokoro-82M model card (retrieved October 23, 2026) |
| Additional source | Model metadata (retrieved October 17, 2026)1234 |
| 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 |