Real-time interpretation, translation that starts as the speaker talks, is moving into production. Meetings need captions, livestreams need dubbing, localization needs bilingual renders, training needs side translation. Vendor numbers are a mess: some quote per-character latency, some first-byte, some just say fast. This review compares five options on engineering usability.
Evaluation criteria and selection
Scope: this piece covers only real-time interpretation and streaming speech translation. Cloud versus local voice cost is a separate cloud versus local voice review. Voice cloning tools are in the voice cloning tools review. The realtime speech model release timeline is in the real-time speech model release piece.
Selection has three rules. First, streaming input, audio in while audio out. Second, at least one side publishes a checkable figure. Third, both cloud and open must be present, so we kept three cloud APIs and two self-hosted open options.
Most important discipline: latency figures are not directly comparable. Every vendor states its metric. Without a shared benchmark this site will not name a fastest. We judge only engineering usability, not translation quality, because no public unified benchmark exists for live interpretation.
The five at a glance
The table below is the overview; details follow per vendor. Every repo number is a GitHub API snapshot dated 2026-09-21. Stars move daily, so treat them as a same-day snapshot.
| Option | Form | Deploy | Languages | License | Snapshot |
|---|---|---|---|---|---|
| Qwen3.8-LiveTranslate | End-to-end SI model | Cloud API | 60 in / 29 out | Closed API | No public repo |
| OpenAI GPT-4o Realtime API | Realtime voice plus gpt-realtime-translate | Cloud API | 70+ in / 13 out | Proprietary | No public repo |
| kyutai-labs/hibiki | Streaming speech translation | Self-host | FR to EN only | Code MIT/Apache, weights CC-BY 4.0 | 1,519 stars / Rust |
| ictnlp/StreamSpeech | Offline plus simultaneous ASR/ST/TTS | Self-host | FR/ES/DE to EN | MIT | 1,291 stars / Python |
| iFlytek simultaneous interpretation | End-to-end SI, SaaS plus API plus on-prem | Cloud SaaS/API, on-prem possible | about 14 to 16 langs | Proprietary | No public repo |
Three are closed cloud APIs, two open self-hosted.
Vendor by vendor
1. Qwen3.8-LiveTranslate
Alibaba Tongyi Qianwen ships a real-time simultaneous interpretation model via the Qianwen AI platform API. Its Interleave architecture weaves audio and text into one sequence, caching heard audio and emitted translation, with a Hybrid MoE Thinker and Talker pair for understanding and voice synthesis that keeps the original voice. Official latency is average per-character latency, LAAL, cut from 2.8 to 2.3 seconds, not end-to-end first-byte. Coverage is 60 input and 29 speech-output languages, two metrics that must not merge. Output is native bilingual same-frame; speaker separation is native via diarization, then per-line voice cloning. Deploy is the Qianwen platform and Alibaba Cloud Bailian WebSocket API, demo at omni.qwen.ai/live-translate. Pricing is not public, follow Qianwen and Bailian docs. It is an API service with no public repo and must not be called open source. Background in the Qwen3.8 SI release hotspot, steps in the Qwen3.8 SI API SOP.
2. OpenAI GPT-4o Realtime API
OpenAI's realtime voice family is an end-to-end multimodal voice-in voice-out model. On 2026-05-07 it shipped a dedicated interpretation model, gpt-realtime-translate, via the realtime translations endpoint, translating spoken input from over 70 languages into 13 output languages without pausing. OpenAI publishes no official end-to-end interpretation latency; a third-party 2026 benchmark measured roughly 711 milliseconds median first translated audio with continuous-speech lag around 3.8 seconds, a benchmark figure not OpenAI's official number and a different ruler from Qwen's LAAL. Output is translated audio plus source and target text, single-language by default, not guaranteed bilingual same-frame. Voice follows the source speaker's tone and pitch and switches per speaker, limited adaptation, not cloning. Deploy is a cloud API over WebSocket or WebRTC, billing 0.034 US dollars per minute, the general series per token. License is proprietary. Timeline in the real-time speech model release piece.
3. kyutai-labs/hibiki
Hibiki is a streaming speech translation model from the French open lab Kyutai, paper arXiv 2502.03382, simultaneous speech to speech with optional voice transfer and text translation. Its Moshi multistream, decoder-only architecture jointly models source and target speech to generate target audio while still processing input. The README gives no end-to-end latency; it emits tokens at a constant 12.5Hz with training up to 120 seconds and a 40-second context window, an engineering constraint not a latency promise. Languages are its weakness: French to English only. Output emits text and audio together, voice fidelity tunable via a CFG coefficient, default 1, empirical 3. Deploy is self-hosted across PyTorch, MLX, MLX-Swift, and Rust, with the small Hibiki-M on phone hardware. License: code MIT for Python and web, Apache-2.0 for Rust, weights CC-BY 4.0 needing attribution. Repo in the hibiki resource page. Snapshot 2026-09-21: 1,519 stars, 119 forks, Rust, last push 2026-09-09.
4. ictnlp/StreamSpeech
StreamSpeech is an All-in-One seamless model from the NLP group at the Institute of Computing Technology, Chinese Academy of Sciences, an ACL 2024 paper. It folds offline and simultaneous ASR, speech translation, and speech synthesis into one two-pass model controlled by CTC alignment to generate while the speaker talks, beating Meta's UnitY on the CVSS benchmark for both modes. Official latency is end-to-end around 320 milliseconds, a CVSS number with S2ST overhead measured by Average Lagging, a different metric from Qwen's LAAL. Languages are Fr-En, Es-En, De-En checkpoints, all to English. Output emits ASR, translation, and synthesis together with word-by-word incremental synthesis. Speaker separation and voice are not strengths; the TTS voice is fixed. Deploy is self-hosted on Python 3.10, PyTorch 2.0.1, fairseq, SimulEval, and a GPU. License is MIT. Snapshot 2026-09-21: 1,291 stars, Python, last push 2025-06-29.
5. iFlytek simultaneous interpretation
iFlytek's simultaneous interpretation product family runs on the Spark speech interpretation model, which the company calls China's first end-to-end speech interpretation model. It exposes three entrances: the consumer app, PC client, and web floating captions; the iFlytek Open Platform simultaneous interpretation API; and an on-premises multi-language meeting system with full-link private deployment. Marketing repeats two numbers, fastest first-character response around 2 seconds for Chinese-English and voice broadcast latency around 2 seconds, both first-response figures not LAAL, so they must not sit beside Qwen's 2.3-second LAAL. Languages are its strong side, about 14 to 16 inter-translatable covering major world languages, wider than hibiki and StreamSpeech but fewer than Qwen and GPT-4o. Output is bilingual captions plus AI audio, with voice cloning since October 2025 at claimed above 90 percent similarity, limited retention not per-speaker cloning. Deploy is cloud SaaS and API, or on-prem keeping data local; consumer captions free then about 0.8 yuan per minute, API from 2000 yuan per 100 hours. License is proprietary, no public repo.
Dimension-by-dimension comparison
The table compresses seven dimensions; the latency column tags each vendor's metric.
| Dimension | Qwen3.8 | GPT-4o Realtime | hibiki | StreamSpeech | iFlytek |
|---|---|---|---|---|---|
| Latency metric and reach | LAAL 2.3s, official | Unpublished official, benchmark first audio ~711ms, lag ~3.8s | Unpublished, 12.5Hz | CVSS ~320ms AL | Marketing first-char ~2s, no official LAAL |
| Languages | 60 in / 29 out | 70+ in / 13 out | FR to EN only | FR/ES/DE to EN | about 14 to 16 langs |
| Output form | Text plus speech, same frame | Speech plus text, single lang | Text plus speech, timed | Multi-result, incremental TTS | Text plus speech, bilingual captions, voice clone |
| Speaker sep and voice | Native, per-speaker | Limited, adapts tone, no clone | Not focus, voice transfer | None, fixed synth | Meeting ASR split, limited voice clone |
| Deploy | Cloud API | Cloud API | Self-host multi | Self-host GPU | Cloud SaaS/API, on-prem |
| Cost | Official docs | From 0.034 USD/min | Compute cost | Compute cost | Consumer from 0.8 yuan/min; API 100h 2000 yuan |
| License and data | Closed API | Proprietary | Code MIT/Apache, weights CC-BY 4.0 | MIT | Proprietary; on-prem keeps data local |
The table's value is not ranking but laying incompatible latency metrics side by side: Qwen 2.3s, StreamSpeech 320ms, iFlytek marketing 2s, and GPT-4o benchmark 711ms measure different things, so do not read them as one ruler.
Scenario ledger
Four common scenes, each a first pick and one-line reason from the metrics above, not a bet on quality.
Cross-border meeting. First pick Qwen3.8-LiveTranslate, same-frame bilingual captions plus native speaker separation in a multi-speaker room, assuming cloud API and data export. If sealed, fall back to iFlytek's private deployment or run hibiki locally, but French to English only.
Cross-border livestream. First pick Qwen3.8, same-frame bilingual plus voice cloning. Tight budget with French source can use hibiki local voice transfer; iFlytek's voice clone works for the host's own voice but trails Qwen and GPT-4o on coverage.
Video localization. First pick StreamSpeech, offline-plus-simultaneous with word-by-word incremental synthesis good for editing, MIT clean for commercial. For light post-captioning, iFlytek's web or client at about 0.48 yuan per minute also works.
Overseas support training. First pick OpenAI gpt-realtime-translate, 70-plus input and 13 output languages with fast Realtime fit. For internal scripts, iFlytek's broad languages plus on-prem build cheaply, but you supply concurrency and compliance review.
Honest limitations
Simultaneous interpretation carries shared weaknesses. First, latency metrics are mostly opaque; cloud quotes first-byte or feel, open source quotes benchmark AL or frame rate, so ranking speed is false. Second, small-language quality is unverifiable beyond a few benchmark tongues, so measure on your own corpus. Third, compliance and data export are hard constraints; closed cloud APIs send audio out while open source and iFlytek keep data local via self-host or on-prem. Fourth, human backup is not optional; keep an interpreter or QA seat, machine interpretation is aid not replacement. For speech-to-speech background see the speech-to-speech resource.
Reminder: all repo stars are GitHub API snapshots dated 2026-09-21 and float daily; licenses follow the repo read that day, recheck before commercial use. These gaps are exactly why this review refuses to crown a single winner.
FAQ
Q1: Why can we not just rank the vendors by latency?
A1: Because each vendor reports a different metric. Qwen's 2.3 seconds is average latency per character, LAAL. StreamSpeech's 320 milliseconds is Average Lagging on CVSS. iFlytek's marketing 2 seconds is first-character response. GPT-4o's benchmark 711 milliseconds is median first translated audio and drifts on continuous speech. The measured objects differ, so lining them up to name a winner is wrong; this review tags each metric and skips ranking.
Q2: Which open-source option covers the widest language set?
A2: Among open options that actually run, hibiki is French to English only and StreamSpeech ships only FR/ES/DE to EN, both narrow. No open option meets wide speech coverage that runs; closed cloud APIs are broader, Qwen 60-in 29-out, GPT-4o 70-plus-in 13-out, iFlytek about 14 to 16.
Q3: Which vendor fits a sealed meeting where data cannot leave?
A3: Prefer options that keep data local. iFlytek's on-prem deployment keeps audio and text on the internal network and can run fully offline. hibiki and StreamSpeech self-host on local GPU. Pure public cloud APIs, Qwen and GPT-4o, send audio offshore, so sealed scenes need a compliance review first.
Q4: Can Qwen3.8 and hibiki run locally, and is source open?
A4: Qwen3.8 is an API service with no public repo, cannot self-host, and must not be called open source. hibiki has code under MIT and Apache-2.0 with weights CC-BY 4.0, self-hosts, and the small model runs on phone hardware, but French to English only. iFlytek cannot self-host source but offers on-prem deployment keeping data local.
Q5: On cost, is cloud API or self-host cheaper?
A5: No universal answer. Cloud bills by usage, OpenAI gpt-realtime-translate from 0.034 USD per minute, iFlytek Open Platform API from 2000 yuan per 100 hours, about 0.2 to 0.33 yuan per minute, Qwen per official docs. Self-host drops API fees but carries compute and ops. Low-volume personal use often favors self-host; high-volume wide-language use favors cloud for peace of mind, but the bill scales with traffic, so measure against your own volume.