Maya ResearchAudio modelOpen weights

Maya 2 Native

Summary

Maya 2 Native is an open text-to-speech model for eleven Indic languages and English. Checkpoint 1.0 was released on October 25, 2026. It generates 24 kHz mono audio in continuous 80 ms frames, offers 20 synthetic voices, and can clone a reference voice from three to ten seconds of audio.

Maya Research reports speech measurements on 6,000 held-out texts, using three ASR judges and a closed-set voice identity classifier. The report gives per-language CER values, failure analyses, streaming latency, and GPU throughput. VoiceArena lists July 2026 as the model release, but gives neither the measurement date nor the exact checkpoint. Its results cannot be tied unambiguously to the October release 1.0.

All benchmark results

Every published result with its source, test setup and comparison models.

Show all 110 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.

bench6k Assamese voice identity, nearest of 20 centroids (Show measurement, test conditions, and source)

Source value: 96.4
Score: 96.4
Metric: closed-set voice identity accuracy
Unit: %
Benchmark version: Maya 2 Native Release 1.0, checked 2026-10-29
Category: voice-identity
Direction: Higher is better
Source type: vendor-reported
Evaluator: Maya Research Maya 2 Native model card
Status: active
Retrieved at: 2026-10-29
Methodology: Vendor-reported result. Maya Research bench6k release evaluation; 6,000 held-out text rows, 500 per language, 31 content categories, 20.3 hours of generated speech; normalization and engine guards on; sampled decoding; one seed per row. Language Assamese (as); nearest-centroid classification among the 20 built-in voices with ECAPA embeddings, using clips of at least one second.

Source Maya Research Maya 2 Native model card

bench6k Bengali voice identity, nearest of 20 centroids (Show measurement, test conditions, and source)

Source value: 98.5
Score: 98.5
Metric: closed-set voice identity accuracy
Unit: %
Benchmark version: Maya 2 Native Release 1.0, checked 2026-10-29
Category: voice-identity
Direction: Higher is better
Source type: vendor-reported
Evaluator: Maya Research Maya 2 Native model card
Status: active
Retrieved at: 2026-10-29
Methodology: Vendor-reported result. Maya Research bench6k release evaluation; 6,000 held-out text rows, 500 per language, 31 content categories, 20.3 hours of generated speech; normalization and engine guards on; sampled decoding; one seed per row. Language Bengali (bn); nearest-centroid classification among the 20 built-in voices with ECAPA embeddings, using clips of at least one second.

Source Maya Research Maya 2 Native model card

bench6k English voice identity, nearest of 20 centroids (Show measurement, test conditions, and source)

Source value: 99.6
Score: 99.6
Metric: closed-set voice identity accuracy
Unit: %
Benchmark version: Maya 2 Native Release 1.0, checked 2026-10-29
Category: voice-identity
Direction: Higher is better
Source type: vendor-reported
Evaluator: Maya Research Maya 2 Native model card
Status: active
Retrieved at: 2026-10-29
Methodology: Vendor-reported result. Maya Research bench6k release evaluation; 6,000 held-out text rows, 500 per language, 31 content categories, 20.3 hours of generated speech; normalization and engine guards on; sampled decoding; one seed per row. Language English (en); nearest-centroid classification among the 20 built-in voices with ECAPA embeddings, using clips of at least one second.

Source Maya Research Maya 2 Native model card

bench6k Gujarati voice identity, nearest of 20 centroids (Show measurement, test conditions, and source)

Source value: 98.9
Score: 98.9
Metric: closed-set voice identity accuracy
Unit: %
Benchmark version: Maya 2 Native Release 1.0, checked 2026-10-29
Category: voice-identity
Direction: Higher is better
Source type: vendor-reported
Evaluator: Maya Research Maya 2 Native model card
Status: active
Retrieved at: 2026-10-29
Methodology: Vendor-reported result. Maya Research bench6k release evaluation; 6,000 held-out text rows, 500 per language, 31 content categories, 20.3 hours of generated speech; normalization and engine guards on; sampled decoding; one seed per row. Language Gujarati (gu); nearest-centroid classification among the 20 built-in voices with ECAPA embeddings, using clips of at least one second.

Source Maya Research Maya 2 Native model card

bench6k Hindi voice identity, nearest of 20 centroids (Show measurement, test conditions, and source)

Source value: 99.6
Score: 99.6
Metric: closed-set voice identity accuracy
Unit: %
Benchmark version: Maya 2 Native Release 1.0, checked 2026-10-29
Category: voice-identity
Direction: Higher is better
Source type: vendor-reported
Evaluator: Maya Research Maya 2 Native model card
Status: active
Retrieved at: 2026-10-29
Methodology: Vendor-reported result. Maya Research bench6k release evaluation; 6,000 held-out text rows, 500 per language, 31 content categories, 20.3 hours of generated speech; normalization and engine guards on; sampled decoding; one seed per row. Language Hindi (hi); nearest-centroid classification among the 20 built-in voices with ECAPA embeddings, using clips of at least one second.

Source Maya Research Maya 2 Native model card

bench6k Kannada voice identity, nearest of 20 centroids (Show measurement, test conditions, and source)

Source value: 97.9
Score: 97.9
Metric: closed-set voice identity accuracy
Unit: %
Benchmark version: Maya 2 Native Release 1.0, checked 2026-10-29
Category: voice-identity
Direction: Higher is better
Source type: vendor-reported
Evaluator: Maya Research Maya 2 Native model card
Status: active
Retrieved at: 2026-10-29
Methodology: Vendor-reported result. Maya Research bench6k release evaluation; 6,000 held-out text rows, 500 per language, 31 content categories, 20.3 hours of generated speech; normalization and engine guards on; sampled decoding; one seed per row. Language Kannada (kn); nearest-centroid classification among the 20 built-in voices with ECAPA embeddings, using clips of at least one second.

Source Maya Research Maya 2 Native model card

bench6k Malayalam voice identity, nearest of 20 centroids (Show measurement, test conditions, and source)

Source value: 99.4
Score: 99.4
Metric: closed-set voice identity accuracy
Unit: %
Benchmark version: Maya 2 Native Release 1.0, checked 2026-10-29
Category: voice-identity
Direction: Higher is better
Source type: vendor-reported
Evaluator: Maya Research Maya 2 Native model card
Status: active
Retrieved at: 2026-10-29
Methodology: Vendor-reported result. Maya Research bench6k release evaluation; 6,000 held-out text rows, 500 per language, 31 content categories, 20.3 hours of generated speech; normalization and engine guards on; sampled decoding; one seed per row. Language Malayalam (ml); nearest-centroid classification among the 20 built-in voices with ECAPA embeddings, using clips of at least one second.

Source Maya Research Maya 2 Native model card

bench6k Marathi voice identity, nearest of 20 centroids (Show measurement, test conditions, and source)

Source value: 99.1
Score: 99.1
Metric: closed-set voice identity accuracy
Unit: %
Benchmark version: Maya 2 Native Release 1.0, checked 2026-10-29
Category: voice-identity
Direction: Higher is better
Source type: vendor-reported
Evaluator: Maya Research Maya 2 Native model card
Status: active
Retrieved at: 2026-10-29
Methodology: Vendor-reported result. Maya Research bench6k release evaluation; 6,000 held-out text rows, 500 per language, 31 content categories, 20.3 hours of generated speech; normalization and engine guards on; sampled decoding; one seed per row. Language Marathi (mr); nearest-centroid classification among the 20 built-in voices with ECAPA embeddings, using clips of at least one second.

Source Maya Research Maya 2 Native model card

bench6k Odia voice identity, nearest of 20 centroids (Show measurement, test conditions, and source)

Source value: 97.4
Score: 97.4
Metric: closed-set voice identity accuracy
Unit: %
Benchmark version: Maya 2 Native Release 1.0, checked 2026-10-29
Category: voice-identity
Direction: Higher is better
Source type: vendor-reported
Evaluator: Maya Research Maya 2 Native model card
Status: active
Retrieved at: 2026-10-29
Methodology: Vendor-reported result. Maya Research bench6k release evaluation; 6,000 held-out text rows, 500 per language, 31 content categories, 20.3 hours of generated speech; normalization and engine guards on; sampled decoding; one seed per row. Language Odia (or); nearest-centroid classification among the 20 built-in voices with ECAPA embeddings, using clips of at least one second.

Source Maya Research Maya 2 Native model card

bench6k Punjabi voice identity, nearest of 20 centroids (Show measurement, test conditions, and source)

Source value: 99.1
Score: 99.1
Metric: closed-set voice identity accuracy
Unit: %
Benchmark version: Maya 2 Native Release 1.0, checked 2026-10-29
Category: voice-identity
Direction: Higher is better
Source type: vendor-reported
Evaluator: Maya Research Maya 2 Native model card
Status: active
Retrieved at: 2026-10-29
Methodology: Vendor-reported result. Maya Research bench6k release evaluation; 6,000 held-out text rows, 500 per language, 31 content categories, 20.3 hours of generated speech; normalization and engine guards on; sampled decoding; one seed per row. Language Punjabi (pa); nearest-centroid classification among the 20 built-in voices with ECAPA embeddings, using clips of at least one second.

Source Maya Research Maya 2 Native model card

bench6k Tamil voice identity, nearest of 20 centroids (Show measurement, test conditions, and source)

Source value: 98.1
Score: 98.1
Metric: closed-set voice identity accuracy
Unit: %
Benchmark version: Maya 2 Native Release 1.0, checked 2026-10-29
Category: voice-identity
Direction: Higher is better
Source type: vendor-reported
Evaluator: Maya Research Maya 2 Native model card
Status: active
Retrieved at: 2026-10-29
Methodology: Vendor-reported result. Maya Research bench6k release evaluation; 6,000 held-out text rows, 500 per language, 31 content categories, 20.3 hours of generated speech; normalization and engine guards on; sampled decoding; one seed per row. Language Tamil (ta); nearest-centroid classification among the 20 built-in voices with ECAPA embeddings, using clips of at least one second.

Source Maya Research Maya 2 Native model card

bench6k Telugu voice identity, nearest of 20 centroids (Show measurement, test conditions, and source)

Source value: 99.4
Score: 99.4
Metric: closed-set voice identity accuracy
Unit: %
Benchmark version: Maya 2 Native Release 1.0, checked 2026-10-29
Category: voice-identity
Direction: Higher is better
Source type: vendor-reported
Evaluator: Maya Research Maya 2 Native model card
Status: active
Retrieved at: 2026-10-29
Methodology: Vendor-reported result. Maya Research bench6k release evaluation; 6,000 held-out text rows, 500 per language, 31 content categories, 20.3 hours of generated speech; normalization and engine guards on; sampled decoding; one seed per row. Language Telugu (te); nearest-centroid classification among the 20 built-in voices with ECAPA embeddings, using clips of at least one second.

Source Maya Research Maya 2 Native model card

Page 1 of 10

All specifications

Architecture

Model class: Open streaming text-to-speech model; Source: Model class
Checkpoint: Version 1.0, released October 25, 2026; Source: Checkpoint

Input and output

Input modalities: Text and optional reference voice; Source: Input modalities
Output modalities: 24 kHz mono audio, streamed in 80 ms frames; Source: Output modalities
Languages: Assamese, Bengali, English, Gujarati, Hindi, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil, and Telugu; Source: Languages

Licence and access

Access: Open weights and inference code; local execution, Python API, OpenAI-compatible HTTP endpoint, and WebSocket streaming; Source: Access
License: Apache-2.0 for the weights and inference code; bundled components retain their own licenses; Source: License

Further details

Voices: 20 synthetic voices; each can speak all twelve model languages, while individual voices have different training-language coverage; Source: Voices
Voice cloning: Zero-shot cloning from three to ten seconds of reference audio. This release does not publish a dedicated voice-cloning benchmark.; Source: Voice cloning
Model size: 2.50B unique parameters; about 5.0 GB of BF16 weights; Source: Model size
Release benchmark: bench6k contains 6,000 held-out texts, 500 per language, 31 content categories, and 20.3 hours of generated speech. The publisher says the texts were excluded from training.; Source: Release benchmark
Streaming throughput: 323 audio seconds per wall-clock second on one H100 NVL with 224 concurrent WebSocket streams. This is a provider measurement for that server run.; Source: Streaming throughput
Benchmark design: The Release 1.0 model card says bench6k contains 6,000 held-out texts, 500 for each of 12 languages, across 31 categories. It produced 20.3 audio hours with normalization and engine guards enabled, sampled decoding, and one seed per row.; Source: Benchmark design
Benchmark judges: The card defines ASR-1 as Maya's primary in-house Indic ASR, Qwen3-ASR as a third judge, and ASR-2 as a second in-house Indic recognizer held out from training, reward generation, and checkpoint selection. Scorer v2 adds Latin/native-script harmonization in code-mixed spans.; Source: Benchmark judges
Benchmark sampling: The release card reports sampled decoding and one seed per row, so the results do not include repeated-seed estimates.; Source: Benchmark sampling
VoiceArena Hindi, Content creation and education: rank: 5 (Historical model release label from VoiceArena is July 2026. Checkpoint identity with Maya 2 Native Release 1.0 (October 25, 2026) is disputed. Published rank range 1-11. The evaluator reports 459 samples; this is not a count of model runs.); Source: VoiceArena Hindi, Content creation and education: rank
VoiceArena Hindi, Overall: rank: 7 (Historical model release label from VoiceArena is July 2026. Checkpoint identity with Maya 2 Native Release 1.0 (October 25, 2026) is disputed. Published rank range 5-9. The evaluator reports 2,268 samples; this is not a count of model runs.); Source: VoiceArena Hindi, Overall: rank
VoiceArena Hindi, Customer support: rank: 7 (Historical model release label from VoiceArena is July 2026. Checkpoint identity with Maya 2 Native Release 1.0 (October 25, 2026) is disputed. Published rank range 5-11. The evaluator reports 817 samples; this is not a count of model runs.); Source: VoiceArena Hindi, Customer support: rank
VoiceArena Hindi, Media and entertainment: rank: 7 (Historical model release label from VoiceArena is July 2026. Checkpoint identity with Maya 2 Native Release 1.0 (October 25, 2026) is disputed. Published rank range 4-10. The evaluator reports 522 samples; this is not a count of model runs.); Source: VoiceArena Hindi, Media and entertainment: rank
VoiceArena Hindi, General conversation: rank: 10 (Historical model release label from VoiceArena is July 2026. Checkpoint identity with Maya 2 Native Release 1.0 (October 25, 2026) is disputed. Published rank range 5-14. The evaluator reports 470 samples; this is not a count of model runs.); Source: VoiceArena Hindi, General conversation: rank

Sources and data date

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 29, 2026. 62 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication.
Unresolved	Human naturalness evaluation Note No controlled human naturalness study or direct TTS comparison is published.
Unresolved	Voice-cloning benchmark Note The model card reports no dedicated zero-shot voice-cloning benchmark.
Unresolved	bench6k confidence intervals Note Maya publishes no bench6k confidence intervals; one seed per row does not estimate repeated-sampling variation.
Unresolved	Voice-identity evaluation sample count Note The per-language rows use 500 benchmark texts each, but the card does not say how many generated clips lasted at least one second and entered the voice-identity calculation. It also gives no exact sample size for the overall mean ECAPA speaker-similarity score.
Unresolved	GPU model for memory table Note The engine documentation gives allocator measurements for these seat configurations, but the reviewed section does not identify a specific GPU model.
Unresolved	VoiceArena checkpoint mapping Note VoiceArena lists July 2026 as the model release, but does not give the measurement date or exact checkpoint revision. Attribution to Maya 2 Native Release 1.0 (October 25, 2026) is unverified.
Not found in the inspected sources	Vals rows in captured payloads Note Nine complete Vals payloads dated 28 October were checked; no exact model ID or unambiguous model label. Limited to these saved benchmarks.
Unresolved	Uncaptured Arena Code/Speech leaderboards Note Text, Agent including Code/Chat/Work, and twelve WebDev views were captured. A separate general Code and Arena Speech leaderboard were not saved as complete tables.
Unresolved	Artificial Analysis catalog reconciliation Note The full current AA model and benchmark catalog was not re-reconciled for this profile in this sidecar. This sidecar makes no AA absence claim; native AA imports remain authoritative.
Unresolved	VoiceArena ci scale methodology Note The methodology page returned no readable text. Raw ci remains in source evidence but is not interpreted as a 95% confidence interval.
Additional source	Maya 2 Native model card and release evaluation (retrieved October 29, 2026) · Maya 2 Native model card and release evaluation · Editorial description reviewed October 29, 2026
Additional source	Maya 2 Native inference code repository (retrieved October 29, 2026) · Maya 2 Native inference code repository · Editorial description reviewed October 29, 2026
Additional source	Maya 2 Native inference engine and serving documentation (retrieved October 29, 2026) · Maya 2 Native inference engine and serving documentation · Editorial description reviewed October 29, 2026
Additional source	VoiceArena (retrieved October 29, 2026) · VoiceArena Hindi TTS leaderboard · Editorial description reviewed October 29, 2026
Additional source	Model metadata (retrieved October 29, 2026)1 2
Additional source	VoiceArena methodology · Editorial description reviewed October 29, 2026