IBM updates Granite Speech 5.0 TurboCTC with seven-language recognition and 470M parameters

IBM has released a multilingual update to Granite Speech 5.0 TurboCTC, extending its 470-million-parameter automatic speech recognition model from English-only transcription to seven languages while preserving the compact Conformer architecture introduced in August 2026.

Announced on October 21, the revision adds French, German, Spanish, Portuguese, Italian and Japanese, alongside native punctuation, capitalization and training improvements intended to increase robustness in noisy and reverberant environments.

IBM reports that average English word error rate improved from 5.78% to 5.51% across nine public short-form evaluation sets, a relative error reduction of approximately 4.7%.

The company additionally reports 19,870 RTFx throughput, compared with approximately 12,600 RTFx at launch. The throughput figure must not be confused with latency for an individual transcription request.

The model retains a 16-layer Conformer, block-wise attention, eightfold temporal downsampling, self-conditioned CTC and non-autoregressive greedy decoding. Its multilingual output vocabulary contains 16,384 BPE units, and training uses approximately 153,000 hours of audio.

Downloadable weights are distributed under Apache 2.0, with support for Hugging Face Transformers and MLX on Apple Silicon. Real-world accuracy and performance remain dependent on language, recording conditions, runtime and hardware.

··········

THE OCTOBER UPDATE EXPANDS THE ORIGINAL ENGLISH MODEL WITHOUT CHANGING ITS CORE ARCHITECTURE.

IBM introduced Granite Speech 5.0 TurboCTC on August 25, 2026, as an English speech recognition model optimized for inference efficiency. The October update maintains the approximately 470-million-parameter encoder-only architecture while expanding multilingual coverage.

IBM did not replace the model with a larger generative speech-language decoder. The substantive changes concern multilingual acoustic training, output vocabulary, transcript formatting and runtime optimization.

The updated model produces punctuation and capitalization directly and IBM reports improved resistance to reverberation and far-field speech. These claims remain subject to deployment-specific evaluation.

........

Property	August release	October update
Parameters	470M	470M
Languages	English	Seven
Architecture	16-layer Conformer	Unchanged
Training audio	~60,000 hours	~153,000 hours
English WER	5.78%	5.51%
Throughput	~12,600 RTFx	19,870 RTFx
License	Apache 2.0	Apache 2.0

........

··········

SEVEN LANGUAGES ARE NOW SUPPORTED THROUGH A SHARED MULTILINGUAL RECOGNITION MODEL.

The updated model supports English, French, German, Spanish, Portuguese, Italian and Japanese through one shared acoustic encoder and multilingual vocabulary, rather than seven separately deployed recognition models.

IBM uses FLEURS, CommonVoice and Multilingual LibriSpeech among its evaluation resources for European languages, while Japanese is assessed using character error rate on CSJ eval1, CommonVoice-17 and FLEURS.

Recognition quality should not be assumed uniform across languages, accents, dialects or recording conditions. Word error rate and character error rate are different metrics and should not be directly compared.

TurboCTC performs automatic speech recognition, not arbitrary speech translation; translation and other generative audio capabilities belong to other Granite Speech models.

........

Language	Code	Recognition
English	en	Supported
French	fr	Supported
German	de	Supported
Spanish	es	Supported
Portuguese	pt	Supported
Italian	it	Supported
Japanese	ja	Supported

........

··········

THE 16-LAYER CONFORMER COMBINES LOCAL ACOUSTIC PROCESSING WITH SELF-ATTENTION.

The Conformer architecture combines convolutional processing of local acoustic features with attention mechanisms capable of modeling longer contextual relationships. Both are useful for recognizing speech, where local phonetic patterns must be interpreted within a wider utterance.

TurboCTC contains 16 encoder blocks with 1,024 hidden dimensions, eight attention heads, 128 dimensions per head and a convolution kernel size of seven.

Block-wise attention uses blocks of 128 frames, restricting attention computation and making sequence processing more predictable. Relative positional embeddings represent temporal relationships between features.

Model size is only one contributor to efficiency; temporal subsampling, block attention and inexpensive CTC decoding also reduce inference work.

........

Component	Configuration
Parameters	~470M
Encoder	16-layer Conformer
Hidden dimension	1,024
Attention heads	8
Head dimension	128
Attention block	128 frames
Convolution kernel	7
Input features	320
Output vocabulary	16,384 BPE
Objective	Self-conditioned CTC

........

··········

TEMPORAL DOWNSAMPLING REDUCES THE ACOUSTIC SEQUENCE BY A FACTOR OF EIGHT.

TurboCTC extracts log-mel spectral features and temporal differences, combining 80 mel features and 80 delta features with adjacent-frame stacking into a 320-dimensional input representation.

Initial features correspond to approximately 100 frames per second. Frame stacking and skipping reduce temporal resolution by two, followed by strided depthwise convolutions and pooled residual connections in the first two Conformer blocks that reduce it by another factor of four.

The resulting encoder processes approximately 12.5 time steps per second, an eightfold reduction. This makes sequence operations less expensive while requiring the trained encoder to preserve enough phonetic information for accurate recognition.

··········

SELF-CONDITIONED CTC SUPPORTS TRANSCRIPTION WITHOUT AN AUTOREGRESSIVE DECODER.

Connectionist Temporal Classification learns to align acoustic representations with target text without explicit frame-level labels for every token. The output sequence is obtained by collapsing repeated predictions and blanks.

TurboCTC uses greedy non-autoregressive decoding rather than generating text tokens sequentially with a separate language-model decoder, avoiding substantial sequential inference overhead.

Self-conditioned CTC projects intermediate recognition predictions back into hidden representations, enabling later encoder layers to refine them. Intermediate and final stages share the output head.

This design favors throughput, although greedy decoding may leave errors that more expensive decoding methods could potentially correct.

··········

THE ENGLISH WORD ERROR RATE IMPROVES FROM 5.78% TO 5.51%.

IBM reports an average English WER of 5.51% on nine public short-form ASR sets, compared with 5.78% for the earlier English-only release.

WER is the sum of substitutions, deletions and insertions divided by the number of reference words. The improvement is 0.27 percentage points, or approximately 4.7% relative error reduction.

The aggregate improvement does not prove that every individual dataset improved, nor does it establish accuracy on specialized terminology, accents, overlapping speech or severely degraded recordings.

Production users should measure recognition error using representative recordings rather than treating a public benchmark average as a deployment guarantee.

........

Metric	Original	Updated
English WER	5.78%	5.51%
Absolute reduction	—	0.27 percentage points
Relative reduction	—	~4.7%
Evaluation sets	9	9

........

··········

IBM REPORTS 19,870 RTFX THROUGH A REUSED COMPILED CONFORMER BLOCK.

IBM reports 19,870 RTFx throughput, compared with approximately 12,600 RTFx for the original release, a ratio of approximately 1.58.

RTFx represents throughput relative to audio duration, not the response latency experienced by an individual user. It can depend strongly on batching, accelerator utilization, audio length, preprocessing and decoding.

IBM attributes part of the gain to compiling one Conformer block and reusing the compiled graph throughout the encoder, reducing compilation overhead and improving execution efficiency.

The October results were reported by IBM while the public leaderboard entry still reflected the earlier English-only release; the new numbers should not be presented as independently confirmed leaderboard standings.

........

Metric	Original	Updated
Throughput	~12,600 RTFx	19,870 RTFx
Relative speed	1.00×	~1.58×
Optimization	Original compilation	Reused compiled Conformer block

........

··········

MULTILINGUAL TRAINING EXPANDS TO APPROXIMATELY 153,000 HOURS OF AUDIO.

The updated model uses approximately 153,000 hours of multilingual audio from publicly available corpora, compared with about 60,000 hours for the original English-only model.

This larger corpus provides exposure to acoustic and linguistic variation, but the aggregate total does not disclose how many hours were used for each language or how recording quality and accents were distributed.

IBM also reports training improvements intended to increase robustness to reverberation and far-field speech. Those claims should be tested independently under representative room acoustics and recording conditions.

··········

NATIVE PUNCTUATION AND CAPITALIZATION REMOVE A SEPARATE PROCESSING STAGE.

The October revision produces punctuated and capitalized transcripts directly, potentially removing the need for a separate restoration model after acoustic recognition.

This can simplify transcription interfaces, subtitles and searchable audio archives by reducing pipeline complexity and an additional potential source of latency.

Punctuation and sentence-boundary quality are not fully characterized by WER. Applications requiring high-stakes transcription should evaluate formatting separately and may still need human review.

··········

THE MODEL IS AVAILABLE UNDER APACHE 2.0 WITH DOWNLOADABLE WEIGHTS.

IBM publishes the exact checkpoint under ibm-granite/granite-speech-5.0-470m-turboctc on Hugging Face, licensed under Apache 2.0.

The license permits use, modification and redistribution, including commercial use, subject to its terms. Downloadable weights enable local inference without requiring every recording to be sent to a proprietary hosted API.

Local execution can support privacy-sensitive workloads, but organizations must still secure recordings, generated transcripts, access permissions and storage. A downloadable model is not a managed service with guaranteed uptime or support.

Users should verify the checkpoint identifier because other Granite Speech variants differ in architecture, capabilities and licensing histories.

··········

HUGGING FACE TRANSFORMERS PROVIDES A DIRECT INFERENCE PATH.

IBM documents native support in Hugging Face Transformers version 5.16.0 and later, using AutoModelForCTC and AutoProcessor.

The model can be loaded with device_map set to auto, and recordings can be processed using the sampling rate expected by the feature extractor. Inference returns token predictions that the processor converts into text.

A minimal installation uses pip install transformers>=5.16.0 datasets. Production applications must additionally manage ingestion, resampling, audio segmentation, validation, concurrency and memory.

The published examples demonstrate checkpoint accessibility, not a complete streaming transcription API or production service.

··········

APPLE SILICON SUPPORT EXTENDS LOCAL DEPLOYMENT OPTIONS.

IBM documents inference using mlx-audio version 0.5.1 or later on Apple Silicon M-series processors. The documented command-line interface accepts a model identifier, audio file and transcript output path.

Apple's unified-memory architecture makes local execution possible without a discrete NVIDIA accelerator, but available memory, processor generation, thermals and runtime configuration affect performance.

The existence of MLX support does not establish identical throughput across M-series devices or universal compatibility with smartphones. IBM does not provide a complete device-by-device memory and latency matrix.

··········

MODEL SIZE DOES NOT DIRECTLY DETERMINE TOTAL MEMORY REQUIREMENTS.

Approximately 470 million parameters require about 940 MB when represented with two bytes per parameter, or 1.88 GB with four bytes per parameter. One-byte storage would represent about 470 MB.

These are arithmetic estimates of parameter storage, not minimum RAM or VRAM requirements. Runtime memory also includes activations, input features, buffers, compilation artifacts and framework allocations.

Longer recordings and concurrent requests may increase memory consumption. Quantization can reduce storage, but actual accuracy and speed depend on the supported implementation.

........

Weight representation	Approximate parameter storage
FP32	1.88 GB
FP16 / BF16	0.94 GB
INT8	0.47 GB

........

··········

TURBOCTC DIFFERS FROM IBM'S LARGER GRANITE SPEECH MODELS.

IBM's October 21 announcements also covered Granite Speech 5.0 2B and Granite Speech 4.1 2B Plus, which address different requirements.

Granite Speech 5.0 2B emphasizes higher-accuracy recognition, speech translation and keyword biasing, while Granite Speech 4.1 2B Plus emphasizes speaker attribution, timestamps and longer multi-speaker recordings.

TurboCTC instead focuses on compact, high-throughput recognition. Capabilities of the larger models must not be attributed to the smaller checkpoint.

IBM's reported 5.06% English WER for Granite Speech 5.0 2B is distinct from the 5.51% reported for TurboCTC.

........

Model	Size	Primary focus
Granite Speech 5.0 TurboCTC	470M	Fast multilingual ASR
Granite Speech 5.0 2B	2B	ASR, translation and keyword biasing
Granite Speech 4.1 2B Plus	2B	Speaker attribution and timestamps

........

··········

BENCHMARK RESULTS REQUIRE CAREFUL INTERPRETATION ACROSS LANGUAGES AND HARDWARE.

The English WER result is an average across selected short-form datasets and cannot determine performance on long recordings, specialized vocabulary or noisy conversational audio.

Multilingual evaluations use different datasets, while Japanese uses character error rate rather than the same word-level metric applied to English.

Throughput measured in a particular accelerator environment cannot be extrapolated directly to laptops or phones. Application-level latency also includes audio ingestion, segmentation, transport and downstream processing.

A meaningful production comparison should measure end-to-end latency, throughput, transcription quality and resource consumption on representative workloads.

··········

THE MODEL'S LIMITATIONS INCLUDE DOMAIN ACCURACY, SPEAKER HANDLING AND LANGUAGE COVERAGE.

TurboCTC is a speech-to-text model rather than a general-purpose audio reasoning, translation or conversational system. Its seven supported languages do not imply uniform accuracy across dialects and recording conditions.

Uncommon names, specialized terminology, overlapping speakers and severe background noise may still cause recognition errors. Greedy CTC decoding is efficient but lacks the search behavior of more elaborate decoders.

The model does not automatically supply all speaker diarization and attribution capabilities found in IBM's larger speech models. Long recordings may require application-level segmentation, and mobile deployment requires device-specific validation.

··········

THE MULTILINGUAL UPDATE STRENGTHENS IBM'S COMPACT SPEECH RECOGNITION OFFERING.

The October update broadens Granite Speech 5.0 TurboCTC from English-only recognition to seven languages while retaining its encoder-only architecture, eightfold subsampling, block-wise attention and greedy CTC decoding.

The combination of native formatting, improved reported English WER, high throughput and downloadable Apache 2.0 weights makes the model relevant for multilingual transcription systems where inference efficiency matters.

The principal remaining questions concern independently reproducible throughput, domain-specific accuracy, multilingual robustness and application-level latency. Production suitability must be assessed on the actual languages, hardware and recording conditions involved.

··········

FOLLOW US FOR MORE.

·····

DATA STUDIOS

·····

datastudios.org