No word boundaries in speech
Spoken Japanese has no word-boundary pauses like space-delimited languages - segmentation comes from grammar and context, not silence. Japanese accuracy is reported as character error rate, not word error rate.
Horizon - Japanese
Transcribe Japanese conversations - from customer support calls to boardroom meetings - with accurate handling of business honorifics, katakana loanwords and real-world audio.
配送のスケジュールを確認したいです。
“I'd like to check the delivery schedule.”かしこまりました。トラッキング番号を教えていただけますか？
“Certainly - could you give me the tracking number?”Language Reality
Most ASR vendors bolt Japanese onto a model built for space-delimited languages. Horizon is trained on how Japanese is actually spoken inbusiness settings - on support calls, in meetings, and in the field - not just read aloud.
Spoken Japanese has no word-boundary pauses like space-delimited languages - segmentation comes from grammar and context, not silence. Japanese accuracy is reported as character error rate, not word error rate.
Many words share pronunciation and differ only by context - 橋 (bridge), 箸 (chopsticks), and 端 (edge) are all read “hashi.”
Business calls shift between formal, polite, and casual registers mid-conversation, changing verbs a model must track in real time. Flattening keigo drops signals QA and compliance teams rely on.
Business Japanese drops English loanwords in katakana - KPI, ミーティング, スケジュール - inside otherwise fully Japanese sentences.
Kansai-ben and other regional accents shift pitch and vocabulary from Tokyo standard, especially on compressed phone audio.
Capabilities
Everything extracted from Japanese audio in a single request.
Sub-500ms first token for live Japanese conversations, across phone lines and web audio.
Handles Japanese-English mid-sentence switching without a separate language-detection pass.
Separates speakers automatically in multi-party calls and meetings.
Japanese punctuation, sentence segmentation and script normalization applied to raw ASR output.
Classify customer intent and detect sentiment directly from the Japanese transcript.
Adapt recognition to your product names, technical terms and industry-specific jargon.
Benchmarks
Horizon tops the OpenASR leaderboard at a 3.10% composite word error rate across eight public benchmarks spanning clean, noisy and multilingual audio, with support across 204+ languages including Japanese.
A dedicated Japanese-language figure isn't published yet - we report benchmark numbers only once we can stand behind them. Request the Japanese benchmark report
#1
OpenASR Leaderboard
3.10%
Composite WER
240+
Concurrent Streams / GPU
204+
Languages Supported
Deployment
01
Managed multi-tenant deployment. Fastest to launch, scales elastically with call volume.
02
Dedicated deployment inside your cloud account. Data never leaves your network perimeter.
03
Runs entirely inside your data center, for regulated workloads with strict residency requirements.
04
Deploy at the branch or device layer for the lowest possible latency and offline resilience.
Data residency: Private VPC and on-prem deployment are relevant for organizations navigating Japan's data protection requirements (APPI) or internal residency policies - Japanese audio and transcripts never need to leave your environment.
Use Cases
Transcribe Japanese customer conversations in real time, power agent assist, and extract sentiment and intent from every call. See contact center deployments.
Generate captions, subtitles and searchable archives for Japanese broadcasts, podcasts and interviews. See media & entertainment.
Transcribe customer interactions, recognize financial terminology, and support compliant recording and audit. See BFSI deployments.
Power Japanese-language customer support with accurate transcription and real-time agent assistance. See telecom deployments.
Integrations
Drop Japanese speech recognition into the voice framework you already use - no separate pipeline to maintain.
from livekit.agents import AgentSession
from livekit.plugins import shunyalabs, silero
# language="ja" pins recognition to Japanese;
# use "auto" for mixed Japanese-English calls
session = AgentSession(
stt=shunyalabs.STT(language="ja"),
vad=silero.VAD.load(),
)Questions
Yes. Horizon Code-Switch recognizes Japanese-English mixed speech natively, without a separate language-detection step or model handoff - useful for business calls where speakers drop in English terms mid-sentence.
Output formatting, punctuation and script conventions can be configured per request through the same smart-formatting layer used across all Horizon languages.
Phone lines, VoIP, web and mobile audio, and uploaded recordings are all supported, in both streaming and batch modes.
Yes. Horizon deploys to cloud, private VPC, on-prem or edge, so Japanese audio and transcripts can stay entirely inside your environment.
Both. Streaming delivers a sub-500ms first token for live conversations; batch is available for recorded audio at scale.
Speech-to-text is billed per minute of audio at the same rate across languages. Current rates are on the pricing page; volume commitments are quoted separately.
Get Started
Whether you're building Japanese voice agents or transcribing conversations at scale, Horizon delivers the accuracy and enterprise reliabilityproduction systems depend on - from pilot rollouts to large-scale Japanese transcription in production.