October shook the interpretation market with one open-weights release. Tencent Hunyuan shipped Hy-SimulMT, weights and stack together, Apache 2.0, deployable on a single H200, official StreamLAAL about 3 seconds. The API camp holds five price lists, from Alibaba's token metering to OpenAI's $0.034 per minute. This review puts all six on one table across speed, accuracy, price, self-hosting, and voice fidelity. One warning: the latency figures span four calibers and cannot be ranked; getting them straight is what makes this comparison worth reading.
A scope note. On September 21 this site ran a five-way realtime interpretation comparison, 37 days ago; this piece rewrites the field: Qwen, OpenAI, and iFlytek stay, joined by October's open-source Hy-SimulMT, Volcano Engine's Doubao Simultaneous Interpretation 2.0, and Google's Gemini 3.5 Live Translate. A separate full-duplex voice comparison covers full-duplex conversation, human and model talking over each other; this piece covers interpretation -- hearing one language, speaking another.
The six, and why these six
Admission criteria: realtime interpretation only (not offline post-processing), with at least one checkable caliber for latency, price, or quality.
The roster: one open-weights self-hosted entry, Hy-SimulMT (Tencent Hunyuan); five commercial APIs, Qwen3.8-LiveTranslate (Alibaba Bailian), Doubao Simultaneous Interpretation 2.0 (Volcano Ark), Gemini 3.5 Live Translate (Google), GPT-Realtime-Translate (OpenAI), and iFlytek Simultaneous Interpretation.
Two roster details. First, Doubao Simultaneous Interpretation 2.0 and ByteDance Seed's Seed-LiveInterpret 2.0 (July 2025) are one model under two names -- product and billing name on Volcano Engine versus model-side name; some articles double-count it, this review does not. Second, no open weights page has appeared for Qwen3.8-LiveTranslate, so its only public form is the Bailian API.
Hy-SimulMT deserves its own paragraph. Open-sourced October 26, it is a 30B-A3B MoE model with roughly 30 billion total parameters, about 3 billion active per inference, shipped in BF16, FP8, GGUF, and an Interleave build that also emits source transcription; the LICENSE file reads Apache-2.0 with a third-party exception list, no additional restrictions. Until now, strong interpretation models were either closed and metered, or open and bottom of the table: Meta's SeamlessStreaming is free and open, yet last in every column of Tencent's four test sets. For the first time the open camp has a contender -- hence this rewrite.
One table, all six
Quality cells use one evaluator's numbers, the four-test-set table in Tencent's Hy-SimulMT README; vendor self-reported figures use different calibers and must not share a column; latency cells carry each vendor's caliber.
| Tool | Form | Price | Latency (caliber matters) | Quality (Tencent-table caliber) | Voice | Self-hosting |
|---|---|---|---|---|---|---|
| Hy-SimulMT | Open weights, self-hosted (Apache 2.0) | Free; anchor BF16 about 57 GiB, one H200 | StreamLAAL about 3s (RealSI long-form; not first-utterance wait) | FLEURS 90.3 / CoVoST2 75.5 / ACL6060 81.2 / RealSI-Sentence 84.7 | Text only; cloning via external TTS (Qwen3-TTS) | Fully self-hostable (SimulInfer) |
| Qwen3.8-LiveTranslate | Bailian API (no open weights) | Audio in 40, text out 100, audio out 160 RMB per M tokens | Character-average LAAL 2.3s (official) | 88.4 / 82.1 / 80.5 / 84.4 | Diarization, voice replication (live, preset) | Not open, no private deployment |
| Doubao Simultaneous Interpretation 2.0 | Volcano Ark API | In 80 / text out 80 / audio out 300 RMB per M tokens; measured about 0.3 RMB/min | Official 2-3s; S2T first character 2.21s, S2S 2.53s | 85.8 / 67.7 / 74.8 / 80.0 | Zero-shot replication, speakers keep their voices | Not open, API only |
| Gemini 3.5 Live Translate | Google API, Translate, Meet | About 0.0368 USD/min effective | Official "a few seconds" | 92.4 / 74.4 / 82.6 / 82.2 | Preserves tone, pace, pitch; no cloning | Not open; Meet base plan caps 90 min/meeting |
| GPT-Realtime-Translate | OpenAI API | 0.034 USD/min | Not published | 77.7 / 50.8 / 76.4 / 76.7 | No cloning found | Not open, API only |
| iFlytek Simultaneous Interpretation | iFlytek Open Platform API | 12-20 RMB/hour; free 20 hours | Not published | 73.4 / 57.0 / 57.5 / 70.9 | Audio narration, no cloning | Private-deployment entry, quote undisclosed |
As one ruler, the quality column shows the field: Gemini takes FLEURS (92.4) and ACL6060 (82.6), Qwen takes CoVoST2 (82.1), Hy-SimulMT takes RealSI-Sentence (84.7). No overall champion. The table comes from Hy-SimulMT's own author; a favorable choice of test sets and scoring model is possible -- reference, not verdict.
Tool by tool
1. Hy-SimulMT: the open-source variable
Official latency is StreamLAAL about 3 seconds on RealSI long-form audio, lowest among compared systems, and officially not the wait for the first spoken word. The team reports GEMBA above the compared open models and four of five commercial APIs (no names given), roughly 85-86 on RealSI long-form, scored by GPT6.1-Sol. A human comparison used 18 recordings of about 6 minutes across 10 domains against 6 certified interpreters; expert lag was 4.5 and 4.7 seconds -- small-scale by Tencent's admission, but a rare human anchor.
Deployment: SimulInfer ships a local stack with web UI, WebSocket API, and optional TTS on port 8080; local verification also found a 30-second per-file cap that rejects rather than truncates. The README is explicit: the base checkpoint returns translated text only, so speech output and voice cloning require a separate TTS, officially Qwen3-TTS-12Hz-0.6B-Base via vLLM-Omni -- reports calling this built-in end-to-end speech are wrong. Tencent's Hy Translate app is a free entry point; scholar endorsements (Wang Weiwei, Zhong Weihe) say "low-risk scenarios."
2. Qwen3.8-LiveTranslate: the cleanest latency caliber
Unveiled at September's Yunqi conference (the Bailian model page separately marks a September 11 listing). Interleave architecture with Hybrid-MoE Thinker-Talker dual modules on a single audio-text stream. It understands 60 languages and speaks 29 (31 text-only). Model ID qwen3.8-livetranslate-flash-realtime, via the Realtime API over WebRTC or WebSocket.
Capabilities include realtime speaker diarization, source and translation in one frame, hot words, and voice replication. Rate limits RPM 10 and TPM 100,000 -- one multi-speaker meeting can hit the ceiling. The latency caliber is the cleanest of the six: officially, 2.3 seconds is character-average LAAL, down from 2.8, against a cited human average of about 4. Integration details: Qwen3.8 live-translate API walkthrough.
3. Doubao Simultaneous Interpretation 2.0: the veteran with two scorecards
Released July 2025, the first to productize end-to-end interpretation. Official figure is 2-3 seconds full-duplex, broken on the same page into S2T first-character 2.21 seconds and S2S output 2.53 seconds. Quality has two coexisting calibers: ByteDance's official RealSI human evaluation gives S2T Chinese-English an average of 74.8 (runner-up baseline 47.3) and S2S 66.3; Tencent's table lists 85.8, 67.7, 74.8, 80.0. Different evaluators, listed side by side. The circulating "GEMBA about 77" appears in no official text and is dropped.
Voice is its signature: zero-shot replication, no sample needed, every speaker keeps their own voice. Postpaid pricing is input 80, text output 80, audio output 300 RMB per million tokens; a 1-billion-token pack costs 56,000 RMB. A third-party developer measured about 1,800 tokens per minute of audio, about 0.3 RMB per minute -- unofficial; cost varies with speech rate.
4. Gemini 3.5 Live Translate: two benchmark wins, vaguest caliber
Released June 9, 2026, end-to-end speech-to-speech translation with automatic detection across 70-plus languages, preserving speaker tone, pace, and pitch, with SynthID watermarking. Official latency is just "a few seconds" -- the vaguest of the six. On Tencent-table caliber it holds FLEURS 92.4 and ACL6060 82.6. Google's pricing page works out to about 0.0368 USD per minute effective; a free tier is listed, quota not spelled out. Product reach is widest: built into Google Translate and Meet, with Meet realtime translation expanding to 63 languages on October 22; base Workspace and AI Pro plans cap interpretation at 90 minutes per meeting, AI Ultra at 8 hours; live streams, recordings, and encrypted meetings unsupported.
5. GPT-Realtime-Translate: cheapest in dollars, bluntest weaknesses
OpenAI's streaming speech-to-speech model on a dedicated v1/realtime/translations endpoint, streaming out translation audio and transcript deltas while source audio is still arriving; 16,000-token context, 2,000-token max output. It takes 70-plus input languages but outputs only 13 (third-party reporting); whether Chinese is among them remains unverified. Price is 0.034 USD per minute, the lowest dollar quote of the five APIs; rate limits tier at 200 to 850 minutes of audio per minute.
Weaknesses, plainly: CoVoST2 of 50.8 on Tencent's table sits near the bottom; no precise latency beyond "keeps pace with the speaker"; no function calling, structured outputs, or fine-tuning; no cloning in public materials.
6. iFlytek Simultaneous Interpretation: lowest entry price, narrowest set
A streaming interface translating audio into multi-language text and audio. The service page says "Chinese-English bidirectional interpretation, currently Chinese-English only," and the API documentation says "Mandarin Chinese to spoken English" -- a narrow set either way. Price is the lowest barrier of the six: a free 20-hour tier (5 channels, 90 days), then 100 hours for 2,000 RMB, 1,000 for 15,000 RMB, 5,000 for 60,000 RMB -- 12-20 RMB per hour. Quality sits in the lower tier on Tencent's table; latency is unpublished. A private-deployment entry exists, quote undisclosed.
Latency: four calibers, no podium
The six latency statements are four calibers plus two blanks: StreamLAAL (a streaming-lag statistic, officially not the wait for the first translated word), character-average LAAL, S2T first-character latency plus S2S output delay, and Gemini's vague "a few seconds"; OpenAI and iFlytek publish nothing. StreamLAAL is a whole-passage lag mean, LAAL a per-character mean, first-character a single instant -- ranking 3, 2.3, and 2.21 together is comparing Celsius with Fahrenheit. All that can be said: the three tools with concrete numbers land between 2 and 3 seconds, overlapping the cited human average of about 4 seconds; no unified benchmark, no conclusion.
Open weights versus the meter: the ledger
Fix the assumptions: a 2-hour meeting daily, 250 working days, 30,000 minutes a year.
Metered APIs: OpenAI, 30,000 x 0.034 = $1,020 per year; Gemini, 30,000 x 0.0368 = about $1,104; iFlytek at the first package's 20 RMB/hour is 10,000 RMB for 500 hours, 12 RMB/hour on the largest package; Doubao at the measured 0.3 RMB per minute is about 9,000 RMB a year (unofficial), thinning to the 56 RMB-per-million pack price at volume; Qwen bills by token, no official per-minute conversion, so no hard number.
Self-hosting: the weights cost nothing, need about 57 GiB of VRAM in BF16, and run on a single H200 (official README anchor, excluding KV cache and runtime overhead). After the one-time hardware outlay, the marginal cost of 30,000 minutes is electricity and maintenance.
The ledger splits three ways. Low volume, a few hundred hours a year: APIs win -- iFlytek has the lowest entry price, OpenAI the lowest dollar rate, no hardware headaches. High volume, one language pair: a year of saved API fees covers a tier of GPU depreciation, and data never leaves the network. High volume, many languages: one H200 model cannot match the 60-to-70-language breadth of APIs -- self-host the dominant pair, route the long tail to APIs.
Cold read
First, no overall quality champion: the benchmark wins split three ways, so match your dominant language and domain material (ACL6060 skews academic). Second, voice capability is uneven: zero-shot cloning only at Doubao; Qwen offers live replication and preset voices; Hy-SimulMT requires building a Qwen3-TTS service; Gemini preserves tone and pace only; OpenAI and iFlytek show none. Third, human backup is not obsolete: interpreters at 4.5 seconds trade wins with every machine here; keep an interpreter or QA seat for formal events.
FAQ
Q1: Why can't the latency figures be ranked directly?
A1: The calibers differ: StreamLAAL is a streaming-lag statistic, LAAL a per-character mean, first-character latency a single instant, and Gemini reports nothing precise. Different quantities -- ranking them together is a category error; this review assigns no podium.
Q2: Hy-SimulMT is free and open -- can it simply replace a commercial API?
A2: Not equivalent. The weights need about 57 GiB of VRAM on one H200-class card; the base model emits text only, so audio requires wiring in Qwen3-TTS; local paths carry a 30-second per-file cap. Light traffic: an API is simpler. High volume, narrow language set: self-hosting pays off.
Q3: Which option keeps audio inside the network?
A3: The only fully open path is self-hosting Hy-SimulMT, where audio never leaves the machine. iFlytek lists a private-deployment entry, quote undisclosed. The other four are public cloud APIs.
Q4: Are Doubao Simultaneous Interpretation 2.0 and Seed-LiveInterpret 2.0 two products?
A4: One product. The latter is ByteDance Seed's model name; the former, its product and billing name on Volcano Engine. Budget for one.
Q5: For a few hours a month, which is cheapest?
A5: iFlytek's free 20-hour tier covers light use; beyond it, packages run 12-20 RMB per hour. Doubao's experience page gives about 100 free minutes daily. OpenAI meters at 0.034 USD per minute. Qwen has no per-minute conversion -- estimate from the token bill.
Related reading