Audio modelOpen weights

Index-Echo-S2ST-2B

Bilibili Index Team

Released: -
Data date: October 7, 2026

Model task: Speech translation with speech output
Access: Public Hugging Face weights for local execution

Index-Echo-S2ST-2B adds a spoken target version to subtitle translation. An S2TT backbone, a Hidden2CV mapper, and CosyVoice3 form its pipeline. The input voice serves as a reference, though Bilibili notes that similarity can vary.

Specifications and access

Specification	Value and source
Model type	Open weightsSource
Model ID	IndexTeam/Index-Echo-S2ST-2BSource
Model task	Speech translation with speech outputSource
Access	Public Hugging Face weights for local executionSource
Input modalities	Chinese or English speechSource
Output modalities	Translated speech conditioned on the input speaker's voiceSource
License	Apache 2.0 according to the Hugging Face model cardSource
Translation directions	Chinese to English, Spanish, or Japanese; English to Chinese, Spanish, or JapaneseSource
Decoder size	2 billion parameters in the text decoder; not the total audio-package sizeSource
Voice conditioning	The target voice is generated using the input speaker as a reference; similarity can varySource
Speech-output architecture	Package combining a frozen S2TT backbone, Hidden2CV mapper, and CosyVoice3 speech generationSource

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.

Index-Echo S2ST WER mean: zh → en (Show measurement, test conditions, and source)

Source value: 0.0709
Score: 0.0709
Metric: end-to-end mean WER content error
Unit: ratio
Benchmark version: arXiv:2609.40181v1, Table 8
Category: audio
Direction: Lower is better
Sample: 100
Source type: vendor-reported
Evaluator: Index-Translate paper v1, Table 8
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Index-Echo-S2ST-2B end-to-end dubbing; Chinese-to-English; 100 retained examples. Paired in-house video dubbing test. E2E content error; shared S2TT text is not scored as S2ST output.
Context: Vendor in-house test without external replication; translation-text judges come from the shared frozen S2TT path and are not attributed to the S2ST model here.

Source Index-Translate paper v1, Table 8

Index-Echo S2ST WER median: zh → en (Show measurement, test conditions, and source)

Source value: 0.029
Score: 0.029
Metric: end-to-end median WER content error
Unit: ratio
Benchmark version: arXiv:2609.40181v1, Table 8
Category: audio
Direction: Lower is better
Sample: 100
Source type: vendor-reported
Evaluator: Index-Translate paper v1, Table 8
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Index-Echo-S2ST-2B end-to-end dubbing; Chinese-to-English; 100 retained examples. Paired in-house video dubbing test. E2E content error; shared S2TT text is not scored as S2ST output.
Context: Vendor in-house test without external replication; translation-text judges come from the shared frozen S2TT path and are not attributed to the S2ST model here.

Source Index-Translate paper v1, Table 8

Index-Echo S2ST source-speaker similarity: zh → en (Show measurement, test conditions, and source)

Source value: 0.743
Score: 0.743
Metric: cosine similarity to source speaker
Unit: ratio
Benchmark version: arXiv:2609.40181v1, Table 8
Category: audio
Direction: Higher is better
Sample: 100
Source type: vendor-reported
Evaluator: Index-Translate paper v1, Table 8
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Index-Echo-S2ST-2B end-to-end dubbing; Chinese-to-English; 100 retained examples. Paired in-house video dubbing test. E2E content error; shared S2TT text is not scored as S2ST output. Speaker similarity is cosine similarity to the input speaker.
Context: Vendor in-house test without external replication; translation-text judges come from the shared frozen S2TT path and are not attributed to the S2ST model here.

Source Index-Translate paper v1, Table 8

Index-Echo S2ST WER mean: zh → es (Show measurement, test conditions, and source)

Source value: 0.0704
Score: 0.0704
Metric: end-to-end mean WER content error
Unit: ratio
Benchmark version: arXiv:2609.40181v1, Table 8
Category: audio
Direction: Lower is better
Sample: 200
Source type: vendor-reported
Evaluator: Index-Translate paper v1, Table 8
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Index-Echo-S2ST-2B end-to-end dubbing; Chinese-to-Spanish; 200 retained examples. Paired in-house video dubbing test. E2E content error; shared S2TT text is not scored as S2ST output.
Context: Vendor in-house test without external replication; translation-text judges come from the shared frozen S2TT path and are not attributed to the S2ST model here.

Source Index-Translate paper v1, Table 8

Index-Echo S2ST WER median: zh → es (Show measurement, test conditions, and source)

Source value: 0.051
Score: 0.051
Metric: end-to-end median WER content error
Unit: ratio
Benchmark version: arXiv:2609.40181v1, Table 8
Category: audio
Direction: Lower is better
Sample: 200
Source type: vendor-reported
Evaluator: Index-Translate paper v1, Table 8
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Index-Echo-S2ST-2B end-to-end dubbing; Chinese-to-Spanish; 200 retained examples. Paired in-house video dubbing test. E2E content error; shared S2TT text is not scored as S2ST output.
Context: Vendor in-house test without external replication; translation-text judges come from the shared frozen S2TT path and are not attributed to the S2ST model here.

Source Index-Translate paper v1, Table 8

Index-Echo S2ST source-speaker similarity: zh → es (Show measurement, test conditions, and source)

Source value: 0.743
Score: 0.743
Metric: cosine similarity to source speaker
Unit: ratio
Benchmark version: arXiv:2609.40181v1, Table 8
Category: audio
Direction: Higher is better
Sample: 200
Source type: vendor-reported
Evaluator: Index-Translate paper v1, Table 8
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Index-Echo-S2ST-2B end-to-end dubbing; Chinese-to-Spanish; 200 retained examples. Paired in-house video dubbing test. E2E content error; shared S2TT text is not scored as S2ST output. Speaker similarity is cosine similarity to the input speaker.
Context: Vendor in-house test without external replication; translation-text judges come from the shared frozen S2TT path and are not attributed to the S2ST model here.

Source Index-Translate paper v1, Table 8

Index-Echo S2ST Kata CER mean: zh → ja (Show measurement, test conditions, and source)

Source value: 0.0493
Score: 0.0493
Metric: end-to-end mean Kata CER content error
Unit: ratio
Benchmark version: arXiv:2609.40181v1, Table 8
Category: audio
Direction: Lower is better
Sample: 200
Source type: vendor-reported
Evaluator: Index-Translate paper v1, Table 8
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Index-Echo-S2ST-2B end-to-end dubbing; Chinese-to-Japanese; 200 retained examples. Paired in-house video dubbing test. E2E content error; shared S2TT text is not scored as S2ST output.
Context: Vendor in-house test without external replication; translation-text judges come from the shared frozen S2TT path and are not attributed to the S2ST model here.

Source Index-Translate paper v1, Table 8

Index-Echo S2ST Kata CER median: zh → ja (Show measurement, test conditions, and source)

Source value: 0.027
Score: 0.027
Metric: end-to-end median Kata CER content error
Unit: ratio
Benchmark version: arXiv:2609.40181v1, Table 8
Category: audio
Direction: Lower is better
Sample: 200
Source type: vendor-reported
Evaluator: Index-Translate paper v1, Table 8
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Index-Echo-S2ST-2B end-to-end dubbing; Chinese-to-Japanese; 200 retained examples. Paired in-house video dubbing test. E2E content error; shared S2TT text is not scored as S2ST output.
Context: Vendor in-house test without external replication; translation-text judges come from the shared frozen S2TT path and are not attributed to the S2ST model here.

Source Index-Translate paper v1, Table 8

Index-Echo S2ST source-speaker similarity: zh → ja (Show measurement, test conditions, and source)

Source value: 0.78
Score: 0.78
Metric: cosine similarity to source speaker
Unit: ratio
Benchmark version: arXiv:2609.40181v1, Table 8
Category: audio
Direction: Higher is better
Sample: 200
Source type: vendor-reported
Evaluator: Index-Translate paper v1, Table 8
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Index-Echo-S2ST-2B end-to-end dubbing; Chinese-to-Japanese; 200 retained examples. Paired in-house video dubbing test. E2E content error; shared S2TT text is not scored as S2ST output. Speaker similarity is cosine similarity to the input speaker.
Context: Vendor in-house test without external replication; translation-text judges come from the shared frozen S2TT path and are not attributed to the S2ST model here.

Source Index-Translate paper v1, Table 8

Index-Echo S2ST CER mean: en → zh (Show measurement, test conditions, and source)

Source value: 0.0372
Score: 0.0372
Metric: end-to-end mean CER content error
Unit: ratio
Benchmark version: arXiv:2609.40181v1, Table 8
Category: audio
Direction: Lower is better
Source type: vendor-reported
Evaluator: Index-Translate paper v1, Table 8
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Index-Echo-S2ST-2B end-to-end dubbing; English-to-Chinese; English-source set began at 200 examples but the paper gives only a 146–198 retained range across directions, not a direction-specific count. Paired in-house video dubbing test. E2E content error; shared S2TT text is not scored as S2ST output.
Context: Vendor in-house test without external replication. The table does not give an exact retained sample count for this English-source direction, so sampleCount is omitted.

Source Index-Translate paper v1, Table 8

Index-Echo S2ST CER median: en → zh (Show measurement, test conditions, and source)

Source value: 0
Score: 0
Metric: end-to-end median CER content error
Unit: ratio
Benchmark version: arXiv:2609.40181v1, Table 8
Category: audio
Direction: Lower is better
Source type: vendor-reported
Evaluator: Index-Translate paper v1, Table 8
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Index-Echo-S2ST-2B end-to-end dubbing; English-to-Chinese; English-source set began at 200 examples but the paper gives only a 146–198 retained range across directions, not a direction-specific count. Paired in-house video dubbing test. E2E content error; shared S2TT text is not scored as S2ST output.
Context: Vendor in-house test without external replication. The table does not give an exact retained sample count for this English-source direction, so sampleCount is omitted.

Source Index-Translate paper v1, Table 8

Index-Echo S2ST source-speaker similarity: en → zh (Show measurement, test conditions, and source)

Source value: 0.623
Score: 0.623
Metric: cosine similarity to source speaker
Unit: ratio
Benchmark version: arXiv:2609.40181v1, Table 8
Category: audio
Direction: Higher is better
Source type: vendor-reported
Evaluator: Index-Translate paper v1, Table 8
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Index-Echo-S2ST-2B end-to-end dubbing; English-to-Chinese; English-source set began at 200 examples but the paper gives only a 146–198 retained range across directions, not a direction-specific count. Paired in-house video dubbing test. E2E content error; shared S2TT text is not scored as S2ST output. Speaker similarity is cosine similarity to the input speaker.
Context: Vendor in-house test without external replication. The table does not give an exact retained sample count for this English-source direction, so sampleCount is omitted.

Source Index-Translate paper v1, Table 8

Page 1 of 2

Sources and data date

Type	Evidence and data date
Research status	Research date October 7, 2026. 1 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication.
Not published	english-source-sample-count Original wording Table 8 gives only 146–198 retained examples across English-source directions, not an exact count for each direction.
Unresolved	text-judge-attribution Original wording Table 8's text-judge column scores the shared frozen S2TT text. It was not attributed to S2ST; a separate text result for the S2ST audio path is absent.
Not found in the inspected sources	independent-s2st-replication Original wording S2ST E2E content errors and speaker similarity come from the model team's in-house test; no external repetition for this package checkpoint was found.
Additional source	Index-Echo-S2ST-2B Hugging Face model card (retrieved October 7, 2026) · Index-Echo-S2ST-2B model card · Editorial description reviewed October 7, 2026
Additional source	Bilibili Index-Translate repository (retrieved October 7, 2026) · Bilibili Index-Translate repository · Editorial description reviewed October 7, 2026
Additional source	Index-Translate paper v1, Table 8 (retrieved October 7, 2026)